Keywords: Autonomous underwater vehicles, Vehicle design, Vehicle performance
Abstract: The interest in extra-large displacement unmanned underwater vehicles (XLD-UUV) is becoming very popular as they offer long rage and large payload capacity and are attracting the attention of large players in the underwater technology domain, as well as a good number of small/medium sized companies. An interesting design trend is being observed: from hull forms based on body of revolution and high slenderness ratio, many designs currently consider different fillet radii. The motivation for this switch appears valid: the manufacturing cost of producing a square cross-sectioned hull with lower filleting curvature; the available internal space can be used far more efficiently; and these shapes can easily rest on and attach to flat surfaces. However, the selection of hull cross-sectional geometry fundamentally influences AUV hydrodynamic drag, at the expense of increased propulsion power and reduced range. This study presents a computational fluid dynamics (CFD) investigation comparing four variations from the initial hull geometry.

Keywords: Autonomous underwater vehicles, Vehicle design, Vehicle performance
Abstract: Autonomous underwater vehicles (AUVs) for seabed operations such as pipeline inspection, coral reef monitoring, and offshore infrastructure assessment require high maneuverability, robustness under disturbances, and computational efficiency on embedded platforms. This paper presents a field-deployable agile AUV equipped with an eight-thruster omnidirectional propulsion system and a hydrodynamically optimized hull, combined with an Adaptive Event-Triggered Nonlinear Model Predictive Control (AET-NMPC) framework. The proposed architecture integrates an NMPC outer loop executed on a Raspberry Pi 4B with a high-frequency PID inner loop running on an STM32 MCU. The controller explicitly considers thruster nonlinearities, saturation constraints, and single-actuator fault reallocation. An adaptive event-triggering mechanism reduces computational burden while preserving tracking performance. Pool experiments under 5 cm/s disturbance currents demonstrate a 60% reduction in RMS tracking error compared with cascaded PID control and a 38% reduction in NMPC execution frequency. Under single-thruster failure, the system maintains graceful degradation with only 7.9% tracking error, validating its robustness for real-world seabed operations.

Keywords: Vehicle design, Autonomous underwater vehicles
Abstract: The Huracan program is exploring the development of a new autonomous underwater vehicle design capable of very long duration autonomous operations, on the order of weeks or more, which may include periods of time in a low power sleep mode. In the first phase of development, a distributed computing and power control system was designed and prototyped, given the critical role that such a system would play in any feasible design. An initial design of the power system was also developed. This paper describes the proposed design as well as initial testing in the laboratory, in the Monterey Bay Aquarium Research Institute’s salt water test tank, and in the field.

Keywords: Autonomous underwater vehicles, Low-cost, scalable oceanographic sensors and platforms, Integration with autonomous platforms and observatories
Abstract: This paper presents an autonomous underwater landing vehicle (AULV) capable of surveying multiple sites in a single deployment, reducing both cost and time. The AULV integrates a variable ballast system for controlled descent and autonomous navigation using GPS, compass data, and a PID controller to travel between waypoints. A previous version of the vehicle was developed for surface navigation and validation studies. An upgraded AULV presented here has been designed for deep-water environments, with an operational depth of up to 300 m. The vehicle first performs autonomous surface navigation, traveling between predefined GPS waypoints using onboard navigation and control algorithms before initiating its descent phase. Once the target location is reached, the AULV begins a controlled descent that requires precise regulation of vehicle weight and orientation. To analyze the vehicle behavior during descent, a dynamic model was developed following a six degree of freedom formulation based on the standard marine craft equations of motion, incorporating rigid body inertia, hydrodynamic damping, and restoring forces due to gravity and buoyancy. The model was implemented includes nonlinear drag forces, buoyancy variation, and horizontal current to estimate trajectory and landing position. Simulation results show that a fully flooded ballast tank produces a terminal descent velocity of approximately 0.87 m/s, which minimizes horizontal drift but results in unsafe landing conditions. A ballast control strategy was introduced in which compressed air is injected during descent to reduce the ballast fraction to approximately 46%, yielding a safe terminal velocity of 0.35 m/s while maintaining acceptable drift. This adjustment is applied at a depth of 175 m during a 200 m descent, balancing landing safety and positional accuracy. Upon reaching the seafloor, proper orientation of the onboard ADCP is critical for accurate measurements, as irregular terrain can induce roll and pitch angles that degrade data quality. A repositioning strategy is implemented in which the vehicle evaluates its attitude after touchdown and, if needed, ascends by modifying its ballast and performs a subsequent descent at a nearby location.

Keywords: Autonomous underwater vehicles, Vehicle performance, Ocean exploration
Abstract: The WHOI Deep Venture class of autonomous underwater vehicles (AUVs) is a compact, lightweight, full-ocean-depth platform designed for flexible, rapidly deployable deep-sea operations. During the 2026 AT51-01 expedition aboard R/V Atlantis, the latest vehicle iteration was evaluated across multiple mission profiles while demonstrating a concept of operations (CONOPS) for supporting Human Occupied Vehicle (HOV) Alvin. Three science missions were completed, including high-resolution coral reef imaging, mesopelagic observations, and pre-dive reconnaissance of a 6,000 m seafloor site. These deployments demonstrated the versatility of the platform in shallow, mid-water, and abyssal environments while validating its ability to provide rapid environmental characterization in support of human-occupied vehicle operations. Survey results from the deep-water mission were delivered within 30 minutes of recovery and directly informed subsequent Alvin dive planning, improving target site selection for scientific observation and sampling. The expedition also demonstrated streamlined launch and recovery procedures, remote operational support, and efficient small-team operations enabled by the vehicle's modular software and hardware architecture. These results highlight the potential of compact, full-ocean-depth AUVs to enhance scientific productivity, operational efficiency, and coordinated robotic–human exploration during future deep-ocean expeditions.

Keywords: Vehicle design, Autonomous underwater vehicles, Hydrodynamics
Abstract: Underwater critical infrastructure, including submarine cables and pipelines, requires regular inspection to ensure reliable operation and environmental safety. Conventional unmanned inspection approaches often rely on separate surface and underwater vehicles, increasing operational and launch-and-recovery complexity. This paper presents the design of a compact hybrid unmanned surface and underwater vehicle (USV–AUV) for near-shore infrastructure inspection. The vehicle is intended to operate primarily at the surface, where satellite-based positioning and communication are available, and to submerge only when close inspection is required. A twin-hull configuration with independently actuated piston-driven buoyancy systems is developed from mission-specific requirements, enabling surface operation, controlled submergence, and underwater inspection within a single platform. The resulting 48 kg design has an operational depth of 25 m, a design depth of 50 m, a target surface speed of 4 kn, and a submerged inspection speed of 2 kn. Analytical resistance calculations and preliminary propulsion sizing indicate that commercially available thrusters provide sufficient thrust margin for the considered operating conditions, although wave-making and twin-hull interference are not included in the surface-resistance model. The study establishes the preliminary mechanical and system architecture of the proposed platform and its suitability for further development toward near-shore inspection missions. Experimental validation of hydrodynamic performance, stability, autonomous mode transitions, localization, and inspection capability remains future work.

Keywords: Vehicle navigation, Autonomous underwater vehicles, Vehicle performance
Abstract: Autonomous Underwater Vehicles (AUVs) are limited by energy capacity. Utilization of path planning algorithms can optimize an AUV path for minimal energy consumption. This paper introduces a modified A* method that includes heading optimization as part of the energy optimal path finding. The hybrid heading strategy allows for two headings to be considered at every grid position. The first heading control method is a constant course control, this is a traditional AUV heading control, the second method is a constant heading control. For a constant heading control, the AUV is not constrained to a track line, instead it drifts offline with the ocean currents, then makes a single corrective maneuver at the end to return to the waypoint. The hybrid A* algorithm in this paper combines both of these heading controls into an energy optimal path. Optimal paths were generated utilizing publicly available ocean current forecast models then verified in two in-water experiments. These in-water experiments predicted 3.2% and 4.3% energy savings. Other scenarios that have not be tested in water have predicted 3-9% energy savings with the hybrid A* algorithm.

Keywords: Vehicle navigation, Real-time decision systems for ocean robotics and autonomous operations
Abstract: Reliable autonomous navigation with an unmanned surface vehicle (USV) in coastal waters requires control strategies that can respond to spatially and temporally varying current fields derived from ocean forecast products. This paper presents a physics-informed reinforcement learning framework for goal-directed marine navigation using realistic surface currents from the Gulf of Maine Operational Forecast System (GoMOFS). We formulate the problem as a continuous-action Markov Decision Process on a 300x300 coastal grid and train the navigation policy using a reinforcement learning (RL) algorithm. To provide flow-aware guidance, we train a streamfunction-based Physics-Informed Neural Network (PINN) on held-out GoMOFS snapshots and use it to generate auxiliary physics features for the policy. We evaluate the resulting method against heuristic and learning-based baselines, including random control, greedy goal-seeking, and other RL-based baselines without physics guidance. On demanding long-range tasks (start--goal separation between [104,278]km), the proposed framework reaches a 68.5% goal-success rate, compared with 44.0% for the Proximal Policy Optimization (PPO) baseline without physics guidance and 0% for the Double Deep Q-Network-based (DDQN) baseline. The physics-informed advantage grows with navigation distance, from +3.4 percentage points at short range (60.4% versus 57.0%) to +24.5 points at long range. Additional experiments show that the relative benefit of the proposed framework increases as current strength grows. These results support using compact physics-informed flow features to improve autonomous marine navigation in realistic, time-varying ocean conditions.

Keywords: Autonomous underwater vehicles, Data assimilation, Real-time decision systems for ocean robotics and autonomous operations
Abstract: The Pacific Infrastructure Supporting Continuous Engineering and Science (PISCES) - Northwest program is developing a multi-modal network of technologies to deliver enhancements to regional-scale forecasting tools and autonomous/unmanned vehicle operations. Oceanographic and atmospheric measurements from stationary profiling moorings, an X-band radar, unmanned underwater vehicles (UUVs), and an autonomous surface vehicle (ASV) were collected in the dynamic coastal ocean and inland waterways of Dabob Bay, WA and Sequim Bay, WA. Data were rapidly processed and standardized through a pipeline that will expose information to regional oceanographic forecasting tools adapted from the Salish Sea and Columbia River Operational Forecasting System (SSCOFS) and a high resolution Regional Ocean Modeling System (ROMS) grid nested inside LiveOcean. Data assimilation techniques being developed for these spatiotemporally-variable measurements are the Data Assimilation Research Testbed (DART) and 4D-Variational Data Assimilation (4D-VAR) framework, respectively. Here, we describe the PISCES system of systems and the individual technologies comprising it in the context of two use cases focused on improving tidal current prediction (Sequim Bay) and sound speed estimation (Dabob Bay). Summary results from multi-modal sensing in both locations are presented from preliminary field events. Ultimately, success for the program will be verified forecast improvements through near real-time data assimilation, demonstrating the value of continuous autonomous sampling of critical coastal waterways.

Keywords: Autonomous underwater vehicles, Imaging and vision
Abstract: Photogrammetry turns overlapping images into a measurable, textured 3D model. The images set an upper bound on the quality of that model: no reconstruction pipeline recovers detail the acquisition did not capture. Each surface must be seen from several viewpoints far enough apart for robust multi-view geometry, while each view stays close enough to the surface normal for its texture to be recovered reliably. In underwater robotics, most autonomous surveys plan for spatial coverage at a constant acquisition rate, and offer no guarantee that either condition is satisfied.

We present a next-best-view (NBV) planning method built on a truncated signed distance field that records, for every surface element, the distinct directions from which it has already been observed. Surface voxels are grouped into oriented surface patches, and candidate viewpoints are generated around them with near-frontal incidence enforced at generation time. Candidates are then scored by an NBV cost function that rewards viewing-direction novelty and parallax diversity against a navigation cost, so that a direction already used is penalised specifically, rather than a surface merely being counted as seen once it has fallen inside the camera field of view.

The method is evaluated in the Stonefish underwater simulator with a BlueROV2-type vehicle carrying a forward stereo pair at a fixed downward pitch, on a simulated isolated object with complex self-occluding geometry, against a reproduced coverage-based planner representative of current underwater practice. It reconstructs the object from 60% fewer images with minor degradation in quality, and acquires them at planned viewpoints only, a mode in which the coverage-based planner fails entirely because geometric visibility alone cannot reveal that a surface still lacks usable viewing directions.

Keywords: Security and resilience of AUV/USV/ROV platforms, Real-time decision systems for ocean robotics and autonomous operations, Autonomous underwater vehicles
Abstract: This work outlines the methodology of an autonomy framework developed for intelligent surfacing planning of unmanned underwater vehicles (UUVs) traversing waterways with high levels of maritime traffic. The surfacing locations are calculated using techniques from prior work that utilize Poisson Point Process methods to optimize the tradeoff between the localization benefits of surfacing with the risk of collision with surface vessels. This prior work is extended from the single agent operations to swarming systems, and merges with swarm surfacing optimization methods that include inter-agent localization and intelligent beacon election algorithms to maximize information gained during surfacing events. This combined approach is applied towards case studies of areas with high maritime traffic and the autonomy validations are conducted in a simulation environment.

Keywords: Automatic control, Numerical modeling and simulation, Real-time decision systems for ocean robotics and autonomous operations
Abstract: Coordinated autonomous marine vehicles can improve ocean observing and forecasting, but strong, time-varying currents make it challenging for teams to maintain prescribed formations and coverage. Using Hamilton–Jacobi reachability and level-set methods, we develop a framework for dynamic coverage maintenance and coordinated formation path planning in complex flows. We determine dynamically feasible coverage regions and compute time-optimal leader-follower paths that preserve and adapt formations. We illustrate these results in double-gyre flows, flows past a cylinder, and realistic ocean currents off Southern California. The findings demonstrate successful dynamic coverage, coordinated square-formation navigation, and deadlock resolution even when ambient currents are stronger than vehicle propulsion.

Keywords: Vehicle navigation, Autonomous underwater vehicles
Abstract: This paper presents a novel initialization method for range-aided simultaneous localization and mapping (RA-SLAM). The general SLAM problem has a well-known separable structure where landmark and robot positions can be solved for in a linear fashion given known robot headings. This paper considers the case where highly accurate heading information is available, which is typical of autonomous underwater vehicle (AUV) navigation, to solve for the range transponder and robot positions. The proposed approach consists of two steps. First, using a generalized trust region subproblem (GTRS), the positions of the transponders relative to the AUV are solved for given the nonlinear range measurements. Second, the relative transponder positions and the known heading of the AUV are used to estimate the transponder and AUV positions by solving a linear least-squares problem. These transponder and AUV position estimates, combined with the highly accurate heading information, provide a reliable initialization method for the general nonlinear RA-SLAM problem. The effectiveness of the proposed approach is tested on a real-world AUV dataset where long baseline (LBL) range measurements are provided in concert with highly accurate heading information provided by an inertial navigation system (INS).

Keywords: Low-cost, scalable oceanographic sensors and platforms, Integration with autonomous platforms and observatories, Autonomous underwater vehicles
Abstract: Collaborative AUV fleets can improve coverage for maritime monitoring, but underwater localization remains difficult without GNSS. Acoustic ranging supports relative positioning through One-Way Travel Time (OWTT) or Two-Way Travel Time (TWTT). OWTT scales better for multi-agent systems because each broadcast yields a range at every receiver, whereas TWTT requires a dedicated round-trip exchange between pairs. Existing OWTT systems typically rely on specialized hardware and highly stable synchronized clocks, limiting their use on small, low-cost platforms.

We extend miniature, low-cost acoustic modems with OWTT capability using a Teensy 4.1 microcontroller, GNSS Pulse-Per-Second (PPS) timing, and a 10 MHz Oven-Controlled Crystal Oscillator (OCXO). Hardware capture timestamps acoustic events against the OCXO, while PPS establishes and disciplines the shared timebase at the surface. A custom 67 x 64 mm integration board combines the microcontroller, GNSS interface, OCXO, and modem connections, allowing the clock to free-run after loss of GNSS during submerged operation.

We validate the system in two field campaigns. In a frozen-lake trial across nine baselines up to 1 km, OWTT and TWTT both achieved centimetre-class repeatability. OWTT remained operational at a 621.86 m baseline where the TWTT return path failed, and returned valid ranges for 84% of transmissions overall compared with 56% for TWTT. In a later open-water harbour trial with four nodes and a moving receiver craft, OWTT matched concurrent RTK GNSS at decimetre RMSE on clear acoustic paths, with 0.15-0.18 m RMSE for the strongest links.

We also evaluate OCXO holdover. After one hour of PPS calibration, the rate-disciplined clock accumulated only 10.8 cm equivalent range error over 48 hours of free-run, remaining within the clear-path acoustic residual. These results show that scalable OWTT ranging can be added to miniature acoustic modems without a chip-scale atomic clock, providing a practical path toward cooperative localization on small AUV fleets. Submerged multi-AUV trials are ongoing.

Keywords: Sound propagation and scattering, Distributed hydrophone arrays and networked acoustic systems, Real-time decision systems for ocean robotics and autonomous operations
Abstract: Short baseline arrays are used to localize and track underwater unmanned vehicles (UUVs) to test and quantify their navigation systems and abilities. Array accuracy relies upon several input factors: the location of the array sensors and general array structure, timing synchronization of the vehicle pinger, and water column sound speed profiles (SSPs). In November of 2025 and May of 2026, a series of large collaborative experiments were performed on the Dabob Range, by NUWC Division, Keyport, University of Washington (UW), and Pacific Northwest National Laboratory (PNNL). During these two experiments a comprehensive set of oceanographic data was collected, utilizing UUVs, a surface unmanned vehicle (SUV), manual conductivity, temperature, and depth (CTD) sensor surface casts, robotic CTD profilers, and moorings. This research evaluates the integration of extensive sound speed profile measurements and models into short baseline tracking arrays and the resulting accuracy dependance for short baseline array localization and tracking. In addition, existing high-fidelity regional oceanographic SSP forecast models are used to compare their range tracking accuracy against real-time collected data.

Keywords: Oceanographic instrumentation and sensors
Abstract: Precise subsea navigation is a requirement for high-resolution seafloor mapping from submerged platforms such as autonomous underwater vehicles (AUVs) or Remotely Operated Vehicles (ROVs). This paper details the design, integration, and field performance of a new, compact navigation system developed by the Monterey Bay Aquarium Research Institute (MBARI) to meet the demands of deep-ocean AUV and ROV seafloor mapping. The system combines two off-the-shelf navigation instruments, integrating a Teledyne RDI Tasman Doppler Velocity Log (DVL) and an Exail Phins C7 Compact Inertial Navigation System (INS). To ensure survivability and performance at extreme depths, both sensors are housed in a custom-designed, in-house manufactured titanium pressure vessel rated to 6,000 meters. A key component of the system is a proprietary MBARI-designed electronics board that provides regulated power to the INS while maintaining electrical isolation between the internal components and from the host vehicle's power bus. This integration streamlines vehicle telemetry and power through a single, unified connector. While the integration of high-sensitivity acoustic and inertial sensors presents inherent challenges regarding electromagnetic interference and acoustic noise, we describe the iterative engineering process used to isolate and mitigate noise issues that initially affected DVL performance such that the instrument was barely usable. The system's reliability and accuracy were validated through an extensive testing regimen, including local sea trials in Monterey Bay and two major scientific expeditions: one in the Beaufort Sea and another spanning the coastal waters of Southern California and Northern Mexico. Field results demonstrate navigational errors are bounded by 0.02% of distance traveled. This represents a substantial improvement in precision over the previously utilized Kearfott SeaDeViL INS, proving the efficacy of this integrated package for high-accuracy survey applications in the deep ocean.

Keywords: Autonomous underwater vehicles, Localization and mapping, Navigation, Underwater docking, Underwater proximity operations
Abstract: Due to the challenges regarding the limits of their endurance and autonomous capabilities, underwater docking for autonomous underwater vehicles (AUVs) has become a topic of interest for many academic and commercial applications. Herein, we take on the problem of relative navigation for the generalized version of the docking operation, which we address as proximity operations. Proximity operations typically involve only two actors, a chaser and a target. We leverage the similarities to proximity operations (prox-ops) from spacecraft robotic missions to frame the diverse docking scenarios with a set of phases the chaser undergoes on the way to its target. We emphasize the versatility on the use of factor graphs as a generalized representation to model the underlying simultaneous trajectory estimation and relative navigation (STERN) problem that arises with any prox-ops scenario, regardless of the sensor suite or the agents' dynamic constraints. To emphasize the flexibility of factor graphs as the modeling foundation for arbitrary underwater prox-ops, we compile a list of state-of-the-art research in the field and represent the different scenario using the same factor graph representation. We detail the procedure required to model, design, and implement factor graph-based estimators by addressing a long-distance acoustic homing scenario of an AUV to a moving mothership using datasets from simulated and real-world deployments; an analysis of these results is provided to shed light on the flexibility and limitations of the dynamic assumptions of the moving target. A description of our front- and back-end is also presented together with a timing breakdown of all processes to show its potential deployment on a real-time system.

Keywords: Sonar imaging, AI/ML for oceanographic and environmental data analysis, Vehicle navigation
Abstract: Reliable low-altitude underwater navigation remains challenging because global positioning is difficult and optical sensing degrades in turbid water. Forward-looking sonar (FLS) and artificial acoustic markers provide a practical alternative, but marker-based localization depends strongly on viewpoint. This work validates an active FLS perception method in which a controller moves toward a fixed goal while preserving or recovering marker geometry. An action-masked Double Deep Q-Network (Double DQN) selects planar translation and pitch actions from a short history of marker features on a precomputed FLS state graph containing 4,830 viewpoints and 26,530 feasible transitions. Checkpoints are selected on 12 validation starts and evaluated on a disjoint set of 12 test starts over three fixed training seeds. The policy reaches the goal in 33 of 36 held-out test rollouts. Among successful rollouts, it reduces the mean number of marker-invisible states from 21.9 to 18.5 and the longest loss interval from 15.1 to 10.9 relative to goal-greedy navigation. It also reacquires markers after 175 of 189 loss events. Under an oracle-correspondence protocol, full 16-keypoint Acoustic-N-point (AnP) support increases from 28.4% to 37.8% of visited states. These results validate the central premise that learned FLS viewpoint control can improve marker observability and geometric support during goal-directed navigation.

Keywords: Benchmark datasets, reproducibility, and open-source frameworks, Marine life and ecosystems, Imaging and vision
Abstract: Coral reef ecosystems are under severe threat from recurrent mass bleaching and climate-driven decline, requiring scalable automated monitoring for effective conservation. Coral species identification depends on labor-intensive manual annotation by experts, which limits large-scale application. This study evaluates DINOv3, a self-supervised vision foundation model, as a frozen feature extractor for automated in situ coral classification. Eight model configurations spanning convolutional networks, Vision Transformers, a Mamba-Transformer backbone, and self-supervised backbones were evaluated on identical data partitions of the GV3 dataset comprising 78,012 in situ coral images extracted from 309 underwater videos and covering 27 species. All partitions enforce strict video-level separation, so that no frame from a source video appears in more than one split. A dual-encoder model that concatenates the representations of a frozen DINOv3 ViT-7B/16 transformer encoder and a frozen DINOv3 ConvNeXt-Large convolutional encoder, followed by a trainable multi-layer perceptron head, attains the highest accuracy on every subset, reaching 91.0% top-1 accuracy on the 27-class task against 82.1% for a previously developed ResNet+ViT model. Residual errors concentrate within a small number of visually similar species groups rather than spreading across the taxonomy, consistent with a combination of intrinsic taxonomic ambiguity and the limited sample size of the rarest classes.

Keywords: AI/ML for oceanographic and environmental data analysis, Imaging and vision, Open-source hardware, software, and community innovation
Abstract: Growth measurements of ex-situ coral specimens are essential for monitoring coral health and treatment outcomes, but current methods are time-consuming, prone to user variability, or inaccurate for irregular morphologies. We present a low-cost system for automated planar morphometric measurement of coral fragments from top-down imagery. The system integrates fine-tuned segmentation, reference-based scale calibration, and geometric correction into a single pipeline to estimate planar surface area and linear dimensions. It is particularly suited for massive and encrusting coral forms, for which planar projection is representative of growth. This pipeline yields an average error of 7.6% against ground truth measurements and predicted segment masks with 97.1% IoU.

Keywords: Imaging and vision, Real-time decision systems for ocean robotics and autonomous operations, AI/ML for oceanographic and environmental data analysis
Abstract: Continuous behavioural monitoring is the obvious use for an aquarium's cameras and remains out of reach: manual ethogram scoring is bounded by expert hours, and prompting a large vision-language model (VLM) on every clip fails on arithmetic rather than on accuracy. The 892 h we hold for a single adult Octopus vulgaris is about 160,000 twenty-second clips, so one five-pass labelling run costs 372 to 1,030 h of wall-clock on an API the aquarium does not have.

We instead treat a 235B-parameter VLM as an offline teacher and distil it into a cascade of small models that runs on site. Cheap gates reject 60% of videos without a full decode, leaving only about 60 h ever decoded; 42.7% of surviving clips still contain no animal, which is why the expensive model is spent last rather than first. Every gate is placed by a measurement on human-labelled data rather than by intuition, and in each case the measurement changed the design.

The students reach 0.665 macro-F1 on a six-class ethogram against a majority baseline of 0.100, and 0.642 mask IoU from 3.2 M parameters with no prompt at inference, above the 0.374 its teacher architecture reaches zero-shot per frame. Applied to the corpus, the pipeline recovers a circadian profile peaking at 45% visible activity at 17:00 against 1 to 5% overnight, and shows human presence nearly doubling movement. Arm-tip speed, computed from the masks without ever seeing the teacher's answer, independently recovers the same ordering of behaviours.

We release 5,222 teacher-labelled clips with full vote distributions, about 25,000 captions, 969 human labels and masks, and three frozen video-level benchmarks.

Keywords: Data visualization, Hydrography / Seafloor mapping / geodesy, Marine life and ecosystems
Abstract: HabiCAT 3D (Habitat Complexity Analysis Tool) was introduced in 2025 as an open-source application for computing 3D rugosity, fractal dimension, and vector dispersion maps on complex habitat models, addressing limitations of 2.5D approaches that fail to capture overlapping features such as branching corals. However, the broader Structure-from-Motion (SfM) photogrammetry workflow generates multiple complementary data products (e.g. estimated camera poses, point clouds, meshes) that ecological analyses typically do not fully utilize. This paper presents substantial extensions to HabiCAT 3D that address these gaps: analysis of point clouds alongside meshes, integration of camera poses and source photos, an experimental VR annotation tool, and volumetric reconstruction of transient objects (e.g. fish, soft corals) lost during SfM reconstruction. Rather than forcing researchers to move between separate 2D tools, HabiCAT 3D unifies all these products within a single, fully 3D environment, where each can be visualized, analyzed, and annotated in the overall context.

Keywords: AI/ML for oceanographic and environmental data analysis, Low-cost, scalable oceanographic sensors and platforms, Fiber-optic distributed sensing (DAS, DTS, DSS)
Abstract: Recent work demonstrates that video-based river stage estimation is operationally feasible using passive optical sensors, edge detection, and machine learning. Under the United States Geological Survey (USGS) Next Generation Water Observing System (NGWOS), the Virginia Institute for Marine Science (VIMS), and the City of Virginia Beach advanced this approach through their participation in the USGS River Image SEnsing (RISE) project and subsequent efforts integrating webcam-derived stage estimation from the Hydrologic Imagery Visualization and Information System (HIVIS). Field deployments typically consist of fixed, high-resolution (4K), infrared-capable cameras across tidal and non-tidal areas across the United States, acquiring imagery at hourly intervals. Image-derived water levels are validated in near real time against co-located Ka-band radar sensors, enabling rigorous performance assessment under operational conditions. This current study focuses on testing, hardening, and scaling the Flood VISION-AI (Flood Visual Inference of Stage Identification by Observation Networks – Artificial Intelligence) algorithm and framework. This computer vision-based system infers water stage from fixed camera imagery, emphasizing transferability, computational efficiency, and integration with existing water level monitoring infrastructure. However, for the broadest applicability of FloodVISION-AI, this system can be adapted to also work with non-fixed view Pan Tilt Zoom (PTZ) cameras. As many web camera models within HIVIS, and other vision-based flood monitoring systems are not necessarily fixed view, this addition is invaluable for providing a product that can be more broadly integrated, but requires additional consideration for both CNN machine learning and image segmentation approaches. To do this most effectively, this paper examines the incorporation of an Image Similarity Index (ISI), and its weighting when identifying whether a web camera is focused on the target area for accurate river stage determination.

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications
Abstract: Seafloor optical imagery provides essential information for marine ecological studies, environmental monitoring, and seabed investigations. As underwater surveys increasingly produce large collections of overlapping images, efficient interpretation of seafloor characteristics over surveyed areas has become an important challenge. However, individual images offer limited spatial coverage, making region-scale analysis challenging and motivating the use of image mosaicking to establish spatial continuity. At the same time, existing semantic image analysis workflows often rely heavily on manual annotation, limiting their scalability and practical applicability. This study proposes a semantic segmentation–based framework for seafloor image analysis that integrates pixel-level classification, image mosaicking, and spatial statistical analysis. Semantic segmentation is employed to perform pixel-level classification of benthic organisms and substrate types, while an image mosaicking is used to establish spatial continuity among individual seafloor observations. In addition, a human-in-the-loop annotation refinement strategy is adopted to improve annotation quality, and a full-factorial hyperparameter analysis is conducted to investigate the effects of key training parameters. Preliminary results show that the proposed model achieves a mean IoU of 62.36%, with strong performance for muddy and sandy substrates and bubble/smoke features, while biological categories and fine boundary regions remain challenging. Under a two-dimensional projection assumption, pixel-based class proportions are further used as spatial statistical indicators to characterize the distribution of seafloor features. The results indicate the feasibility of integrating semantic segmentation with image mosaicking for quantitative spatial analysis of seafloor imagery. The proposed framework provides a reproducible workflow for transforming image data into spatially interpretable information, supporting large-scale seafloor analysis.

Keywords: Passive observing sensors, Source localization, Distributed systems, Multi-array signal processing
Abstract: Distributed multi-array passive source localization often suffers from severe ambiguity in multisource environments, because bearing lines from different sources can intersect to form spurious sources and ghost regions in localization maps. To mitigate this problem, this paper proposes a correlation-based passive source localization (CBPSL) framework that exploits the temporal and spatial consistency of source-specific frequency patterns in frequency–azimuth (FRAZ) representations. Temporal consistency is evaluated between adjacent snapshots within each array, while spatial consistency is evaluated across arrays at the same snapshot. These correlation measures are combined to construct a weighting map, which is applied to the triangulation-based beam power map to emphasize true source positions and suppress spurious ones. Controlled simulations are first used to isolate the individual effects of noise level, spectral overlap, and the number of sources on CBPSL performance. CBPSL suppresses ambiguity more effectively than conventional triangulation, and is more robust to noise than a recently proposed reference method. Realistic simulations are then conducted using a normal mode propagation model that includes low-frequency dispersive propagation, Doppler shifts, and bottom-reflected multipath. Two environments are considered: a constant sound speed profile (SSP) with direct-path arrivals, and a refractive SSP dominated by bottom-reflected arrivals. A conic-angle-induced localization error emerges near the end-fire region of a nearby array, more pronounced under the refractive SSP, but the estimates recover near the true trajectory once the source leaves this geometry. Experimental validation using the SWellEx-96 dataset, including one- and two-source cases, further confirms that CBPSL suppresses ambiguity and provides clearer localization than triangulation.

Keywords: Search-and-rescue technologies and emergency response, Surveillance, detection, and threat-assessment techniques, Model-based signal processing techniques
Abstract: Locating Flight Data Recorders (FDRs) in deep-water environments remains a critical challenge, as conventional search methods—relying on zig-zag sonar imaging or acoustic amplitude trends—are often confounded by rough seafloor topography and multi-path interference. This study proposes a novel localization approach by drawing a mathematical analogy between Underwater Locator Beacon (ULB) signals and seismic wave propagation. While seismic networks utilize stationary sensors to locate a single event, our method employs a single GPS-equipped hydrophone drifter to capture repetitive 1 Hz ULB pings while moving with ocean currents. By treating the drifter’s sequential positions as a "quasi-static" sensor network and correcting for the fixed 1 Hz repetition rate, we adapt Geiger’s Method to estimate the ULB’s seafloor coordinates (x, y, z) and epoch time (t). The signal processing architecture utilizes a heterodyne receiver design to optimize detection. Received pings are isolated via a high-Q bandpass filter at 37.5 kHz and mixed with a 35 kHz Local Oscillator (LO). This produces a lower sideband at 2.5 kHz and an upper sideband at 72.5 kHz; the latter is suppressed via a 3 kHz low-pass filter. The retained 2.5 kHz target signal is processed using the Goertzel algorithm for precise time-of-arrival (TOA) detection. These arrival times are synchronized with GPS coordinates and processed through a sparse subset of detected pings to enhance computational efficiency. Validation was conducted through numerical simulations and field testing. Simulations demonstrate that the method remains effective at ocean drift velocities between 0.5 and 6 knots. Preliminary short-range tests in a port environment successfully localized a ULB within 2 meters of accuracy from a distance of 100 meters, utilizing only two minutes of recorded data. This approach offers a mathematically rigorous and hardware-efficient alternative to traditional maritime search procedures, significantly reducing the uncertainty radius for secondary recovery missions.

Keywords: Passive observing sensors, Sonar signal processing, Signal detection, system-informed neural network, uncertainty quantification
Abstract: Accurate and high-resolution frequency estimation is essential for underwater surveillance with passive sonar systems. Recently, system-informed neural networks (SINNs) have been introduced to address

the limitations of conventional methods by providing improved generalization, fast inference, and outputs consistent with system models. Prior work has extended SINN with multiple measurements to leverage consistency across them, yet some frequency components remain weak or masked by noise. In this study, we enhance SINN performance using a deep ensemble with uncertainty quantification (UQ), leveraging the temporal continuity of frequency components. Multiple SINNs are trained with distinct initializations on separate training sets, producing diverse predictions for the same test sample. The resulting UQ provides insights into the prediction uncertainty associated with specific frequency bin components. Consequently, we unify adjacent UQs in time to refine the estimated frequency spectrum by further suppressing residual noise. Simulations demonstrate improved noise reduction, though at the cost of occasionally missing minor components.

In-situ evaluations using two acoustic data confirm that the proposed scheme surpasses nominal SINN and approaches sparse recovery methods, achieving clearer detection of major components under low signal-to-noise ratio, long-range conditions, albeit at the cost of weakened estimation of minor components.

Keywords: Passive and active acoustic monitoring, Model-based signal processing techniques, Sonar signal processing
Abstract: Estimating ship speed and closest point of approach (CPA) range from passive acoustic signals depends strongly on the acoustic representation used for motion characterization. This study presents a controlled comparison of LOFAR and cepstrum representations generated from identical simulated acoustic scenes using the same Bayesian multi-task estimation framework. LOFAR emphasizes tonal and Doppler-related spectral variations, whereas cepstrum emphasizes delay-related structures produced by multipath propagation. The framework jointly estimates ship speed and CPA range while quantifying predictive uncertainty and providing auxiliary information on input image quality. Overall, LOFAR achieved lower speed and CPA range RMSEs than cepstrum, with 1.99 versus 2.81 m/s and 58.91 versus 82.70 m, respectively. Under clean conditions, however, the two representations showed comparable estimation accuracy. As acoustic noise increased, cepstrum degraded more rapidly, accompanied by a substantially larger increase in aleatoric uncertainty. The two representations also showed similar scene-dependent correlation patterns between speed and CPA range estimates under clean conditions, but this structure was less stable for cepstrum as noise increased.

Keywords: Array signal processing and array design, Passive observing sensors, Marine life and ecosystems
Abstract: A framework for automatic underwater acoustic signal detection with direction finding capabilities employing a 192-element coherent hydrophone array is presented. Beamformed time series are analyzed via two complementary methods: the mean power spectral density (PSD) that reveals energetic arrivals, and the cyclic spectral coherence (CSC), estimated with the fast spectral correlation algorithm, that can be applied to temporally patterned signals. Both quantities are formed into bearing–time images, converted to bearing with respect to true north, and processed by a two-dimensional cell-averaging constant false alarm rate (CA-CFAR) detector that adapts the threshold to the local background in bearing and time; the detections are reduced to one point per source and time unit, yielding bearing-time trajectories for multiple sources. The framework is applied to data recorded in the Gulf of America on 6 April 2024 using the Northeastern University inhouse developed array system. Both underwater acoustic processing methods are tested by applying them to analyze Rice’s whale low frequency (LF) call sequences present in the array dataset. The CSC spectrum, averaged along the detected trajectory, exhibits a harmonic series whose fundamental gives an inter-pulse interval (IPI) of 0.63–0.65 s for the Rice’s whale LF calls. The Graphics Processing Unit (GPU) implementation processes the data faster than real time, making the framework suitable for onboard passive acoustic monitoring.

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Benchmark datasets, reproducibility, and open-source frameworks, Real-time decision systems for ocean robotics and autonomous operations
Abstract: Accurate underwater navigation is challenging due to the unavailability of GNSS signals and time-varying uncertainties in inertial and Doppler velocity log measurements. Several works have investigated the contribution of neural-aided adaptive process noise covariance to positioning accuracy, yet without a common ground for comparison and analysis. To bridge this gap, we benchmark three nonlinear Kalman filter variants under an adaptive process noise estimation mechanism designed using a Sim2Real training approach. Through this comparative benchmarking framework and using 80.8 minutes of real-world autonomous underwater vehicles trajectories, we demonstrate that the Sim2Real adaptive invariant Kalman filter outperforms the baseline Sim2Real adaptive extended Kalman filter by 57.2%. These results provide practical guidance for selecting a nonlinear filtering approach in Sim2Real strategies, and highlight the benefit of combining adaptive process noise estimation with invariant filtering for underwater navigation.

Keywords: MEMS-based sensing technologies and miniaturization, Oceanographic instrumentation and sensors, Smart cables and telecom-enabled ocean observation
Abstract: Deep ocean pressure monitoring is important for detecting tsunami propagation and small seafloor pressure changes associated with crustal deformation. Although instrumental drift in quartz crystal gauges has been well studied, the drift characteristics of silicon resonant pressure gauges recently adopted in the Nankai Trough seafloor cable network (N-net) have been evaluated only to a limited extent. We evaluated the drift characteristics of the same type of silicon resonant pressure gauges used in N-net and examined whether a differential operation analogous to the A-0-A method could suppress common instrumental drift. Laboratory experiments were conducted at 50 MPa for 84 days under stable temperature conditions. The results show that the dominant exponential time constants are approximately 4–6 days, that differential operation reduces the linear drift to less than 1 hPa/year under the tested conditions, and that the normalized inter-unit variability show similar temporal behavior among the tested pressure modes. These results support precise long-term seafloor pressure monitoring with the silicon resonant pressure gauges.

Keywords: Systems and observatories, Passive observing sensors, Seafloor pressure observation, Tsunami disaster mitigation, Operational observation
Abstract: Seafloor pressure observations mainly use precise quartz resonant sensors at the present due to its high sensitivity. Recently, a new silicon resonant sensor using Micro Electro Mechanical Systems (MEMS) technology has emerged with low-power consumption, compact-size, constant quality, and high sensitivity. High stability is expected due to small size of a silicon resonant pressure sensor, and individual difference of each sensors is believed to be small because of using the MEMS technology. Influence of movement of the sensor itself is small due to mechanism of pressure transmission. Nankai trough seafloor observation network for earthquakes and tsunamis named N-net is equipped with the new sensors. Before routine observations using this new sensor, we evaluate performance of the new silicon resonant sensor for seafloor pressure observation. An existing free-fall pop-up type ocean bottom pressure gauge has a quartz crystal resonant sensor. We replaced the sensor to the new silicon resonant pressure sensor, and performed seafloor observations off Boso peninsula near Tokyo. We obtained pressure data from two new sensors. Noise spectra of the new sensor was compared to that of the ordinary crystal sensor which was installed at the same position. The new and ordinary gauges have identical noise level at periods longer than about 50 s. Although the new silicon gauges seems to have slightly higher noise levels at periods between 50 s and 10 s, the new pressure gauge can detect pressure change larger than 1 hPa through the whole observation band. The new pressure gauge observed sea level fluctuation by volcanic eruption in Tonga and the identical waveforms were obtained from the new and ordinary pressure gauges.

Keywords: Oceanographic instrumentation and sensors, Low-cost, scalable oceanographic sensors and platforms, Systems and observatories
Abstract: The NOAA Center for Operational Oceanographic Products and Services (CO-OPS) maintains the National Water Level Observation Network (NWLON) as the nation's authoritative source for water level data. As development and interest in lower-cost, simplified water level systems grow, evaluating their performance against established NWLON standards is critical. This study presents a comparative performance analysis of the Hohonu ultrasonic sensor and VegaPuls C22 radar sensor against the Xylem/YSI radar and data collection platform system currently used by CO-OPS across NWLON. Through laboratory and field testing at three diverse NWLON sites, we evaluate the performance, operational feasibility, and reliability of these technologies. Our results assess the feasibility of incorporating alternative sensors into NWLON, their suitability for non-NWLON applications, and the viability of integrating third-party data into CO-OPS products and services, including the Coastal Ocean Reanalysis (CORA), VDatum, and the Coastal Inundation Dashboard.

Keywords: Buoy Technology, Oceanographic instrumentation and sensors, Meteorology
Abstract: The Tropical Atmosphere Ocean (TAO) array consists of 48 moored buoys in the tropical Pacific Ocean, which provide meteorological and oceanographic observations for the purpose of improving ENSO analysis and forecasting. In recent years, the satellite community noticed a systematic high bias of approximately 0.5-0.8 m/s in the TAO wind speed measurements compared to satellite observations, which began abruptly in mid-2020. The National Data Buoy Center (NDBC) launched an investigation and determined that the bias corresponds to an upgrade to the wind tunnel used to calibrate the TAO anemometers.

NDBC has partnered with the Pacific Marine Environmental Laboratory, the National Weather Service Sterling Field Support Center, and the National Institute of Standards and Technology (NIST) to compare calibration biases from multiple wind tunnels. An expected outcome of this partnership is to validate the NDBC wind tunnel calibrations against a NIST-traceable standard and determine if a bias correction needs to be applied to the TAO wind speed data. In addition, this partnership will work toward ensuring a consistent, NIST-traceable calibration framework throughout the National Oceanic and Atmospheric Administration (NOAA).

TAO wind observations remain operationally available and scientifically useful; however, users are cautioned that the wind speeds may be biased by up to +0.8 m/s for buoys deployed between 2020 and 2025, which may contaminate analyses of long-term trends and variability. Upon completion of the calibration assessment, NDBC will communicate final recommendations to the scientific community and apply any necessary bias corrections to the TAO wind speeds.

Keywords: Marine optics technology and instrumentation, Oceanographic instrumentation and sensors, Buoy Technology
Abstract: Reconfigurable Ultraviolet Beam Illuminator Controls Biofouling in Salinity Sensors

Biofouling degrades the accuracy of precision salinity sensors during extended ocean deployments. Chemical antifoulants such as tributyltin (TBT) are effective but toxic, costly, and banned in some jurisdictions. Although short wavelength ultraviolet (UV) light has been shown to control biofouling in oceanographic instrumentation, conventional UV illuminators do not deliver sufficient optical intensity to the far end of the elongated conductivity cells used in many commonly deployed salinity sensors.

A dual-lens UV-C (5 mW) beam illuminator has been developed to generate an intensity up to 2.4 mW/cm2 at the far end of a Sea-Bird 37 conductivity cell, which exceeds the level needed to control biofouling by more than a factor of 20. In two offshore deployments, a salinity sensor protected by a UV beam illuminator was compared with identical sensors first unprotected and then protected by TBT. The UV-protected sensor easily exceeded the performance of the unprotected sensor. Then the UV-protected sensor matched the performance of the TBT-protected sensor, indicating sufficient optical intensity for biofouling control throughout the conductivity cell. The region around the entrance of the conductivity cell was kept clean as well. This beam illuminator can alternatively be configured to protect shorter, larger-diameter conductivity cells, delivering even higher optical intensities over larger areas. The intensity levels delivered by this device should enable significant duty cycle reduction, eliminating the need for continuous operation of this 2 W illuminator.

Keywords: Pollution monitoring, Ocean color and hyperspectral measurements, Coastal zone management
Abstract: Global monitoring of coastal eutrophication under Sustainable Development Goal (SDG) Target 14.1 requires scalable, scientifically robust frameworks that bridge global consistency with regional accuracy. While global ocean color products provide essential baseline monitoring, they frequently fail in nearshore waters due to land-adjacency effects and coarse spatial resolution. This study presents the technical operationalization of the Global Eutrophication Watch (GEW) framework, integrating a multi-scale satellite approach: ESA OC-CCI 1-km merged datasets for decadal trend analysis across 764 global coastal segments and JAXA GCOM-C/SGLI 250-m high-resolution observations for national-scale monitoring in Japanese waters. Rigorous quantitative validation against long-term in situ chlorophyll-a measurements (2016–2023) in Nanao Bay demonstrates the robust performance and technical feasibility of the proposed framework for operational marine assessment.

Keywords: Pollution monitoring, Data visualization
Abstract: This work enhances recent graph-based representations of compound peaks in GC-MS (gas chromatography - mass spectrometry) and similar analytical hardware to quantify the degradation of contaminants in marine pollution studies. Of specific interest are degradation profiles of hydrocarbons in crude oil from oil spills, a major source of oceanic and coastal pollution. Accordingly, we leverage recent work on persistence graphs for raw signal analysis of the GC-MS signals across a portfolio of dated crude oil samples collected from the 2010 Deepwater Horizon oil spill disaster. In this work, we specifically present persistence graph visualization of ions Benzo (a) Anthracene 228 and C-2 dibenzothiophene 212 as examples of ion profiles that are typically studied in the degradation of crude oil.

Keywords: Oceanography: physical, geological, chemical, biological, Hydrodynamics, Microbial community structure and function
Abstract: Marine biofilms can lead to functional and financial problems for the shipping industry. The application of ultraviolet light (UVC) is of growing interest for biofilm prevention. Most of the current research on UVC in the marine environment has focused on growth under static conditions, not taking into consideration the different communities and structure of biofilms which form under flow regimes. To address this gap in knowledge, a field experiment was designed to examine how biofilm which forms under flow can be reduced by the application of UVC and how these changes manifest in the mitigation of drag penalties for ships. Field biofilms were allowed to accumulate in an open grow channel on epoxy-coated panels. After several weeks, the biofilm was assessed and the panels were then placed in a separate field deployed water tunnel to measure drag using a streamwise pressure gradient. Panels were then placed back in the grow channel for UVC treatment. The biofilm assessment and drag measurements were repeated after the UVC treatment. A reduction in the biofilm viability, significant declines in chlorophyll a, as well as changes in bacterial and algal communities were found post UVC treatment. There was a decrease in the drag coefficient after UVC treatment with reductions comparable to a smooth acrylic reference panel. These results demonstrate that UVC can reduce biofilm formation, which results in lower drag penalties

Keywords: Pollution clean-up and pollution remediation, Pollution monitoring, Numerical modeling and simulation
Abstract: Marine microplastic (MP) pollution poses a severe threat to aquatic ecosystems. While magnetic nanoparticle adsorption can effectively aggregate MPs, these delicate flocs easily fracture under high-velocity shear, necessitating low-speed hydrocyclone separation. However, hydrodynamic behavior and internal flow structures in this low-speed regime remain poorly understood. This study uses OpenFOAM v2012 to numerically evaluate the combined effects of wall geometry and nine vortex-generating internal configurations (varying upper diameter, and placement height) on MP recovery and flow dynamics. The optimal performance (97.9% underflow recovery) has been achieved via geometric optimization. Velocity evaluations reveal that a smaller radial distance near the bottom locally amplifies centrifugal acceleration and suppresses upward axial drag without altering global tangential velocity profiles. Conversely, internal vortex generators universally impaired separation efficiency; only one configuration exceeded 90% recovery. Although internals elevated upper-chamber tangential velocity, their top physical edges induced localized upward axial velocity jets, creating short-circuit currents that swept particles into the overflow. Placement height exerted a stronger influence than diameter, with hydrodynamic disruptions spatially confined to the near-field wake. These findings demonstrate that suppressing localized upward axial drag is far more vital for low-speed aggregate recovery than global swirl amplification, providing actionable design guidelines for industrial hydrocyclones.

Keywords: Numerical modeling and simulation, Data fusion, uncertainty quantification, and predictive modeling, Marine GIS and data fusion
Abstract: This study details the development, implementation, and validation of a fine-scale sub-grid hydrodynamic modeling framework designed to simulate coastal inundation at high spatial resolution for Charleston, South Carolina, and Savannah, Georgia. The work builds on prior large-scale storm tide simulations using the open source SCHISM hydrodynamic model, extending those capabilities to resolve localized, street-level flooding dynamics through the application of the UnTRIM3 sub-grid model. The primary objective is to improve predictive fidelity for three key inundation characteristics: timing, maximum spatial extent, and depth, all of which are critical for coastal risk assessment and infrastructure planning. The UnTRIM3 sub-grid approach integrates high-resolution bathymetric and topographic data derived from lidar and sonar into a comparatively coarse computational mesh.

The Charleston and Savannah study domains illustrate the scalability of the approach. In Charleston, a base grid with approximately 55,400 cells is augmented by over 1.7 million sub-grid cells at resolutions as fine as 1x1 m, sufficient to resolve individual buildings. Savannah’s domain is even larger, with over 12.6 million sub-grid cells nested within approximately 181,903 base cells. This multi-scale discretization enables detailed representation of hydraulic connectivity pathways that are typically unresolved in traditional models, such as flow through narrow streets or around urban structures. The street-level model was validated against historical storm events, beginning with 1989 Hurricane Hugo in Charleston. Boundary conditions include NOAA tide gauge data, USGS river discharge measurements, and meteorological forcing from the National Weather Service. A three-day simulation achieves strong agreement with observed water levels, with R²=0.9702 and an RMSE=6.42 cm. The model also quantifies infrastructure impacts, estimating over 5,400 structures experiencing less than 1 ft. of flooding and total economic damages of approximately 1.24 billion (2026 USD equivalent).

Keywords: Protection of offshore energy and critical maritime infrastructure, Energy-harvesting and long-duration monitoring solutions, Ocean energy
Abstract: Coastal power plants are essential infrastructure for meeting energy demands, yet they face increasing challenges from environmental threats, including seasonal sargassum accumulation events that can obstruct water intake systems. Sargassum poses significant operational challenges by clogging intake valves and disrupting cooling systems, reducing plant performance and threatening the energy resilience of coastal communities. This paper addresses sargassum accumulation as a design-based threat (DBT) using the St. Croix facilities in the U.S. Virgin Islands as a case study. Utilizing the Microgrid Design Toolkit (MDT), this study analyzes satellite-derived data to characterize sargassum occurrence and biomass density, and develops fragility curves to quantify power plant vulnerability. This work provides a framework to enhance the resilience of coastal energy infrastructure.

Keywords: Protection of offshore energy and critical maritime infrastructure, Energy-harvesting and long-duration monitoring solutions, Ocean energy
Abstract: Sargassum is a buoyant brown algae in the Atlantic that forms large mats impacting coastal regions. These blooms can obstruct seawater intakes at power and desalination plants, reducing efficiency and increasing the risk of shutdowns. This study presents a CFD framework in OpenFOAM to simulate interactions among waves, intake flow, and sargassum. The model couples a multiphase solver with a Lagrangian approach to represent transport and accumulation. Results show aggregation near intake zones driven by wave and suction effects. Future work will focus on quantitative validation and integration with power system models to assess performance impacts.

Keywords: Numerical modeling and simulation, Ocean energy, Protection of offshore energy and critical maritime infrastructure
Abstract: Sargassum accumulation events in St. Croix have become a growing concern for coastal power plants, as floating biomass can obstruct seawater intake systems and reduce the cooling-water available for plant operation. This reduction in cooling capacity can affect thermodynamic performance and ultimately impact power generation. This paper presents a multi- domain modeling framework of a combined-cycle power plant to evaluate the thermo-electric effects of cooling-water reduction caused by sargassum intrusion. The model includes the Brayton cycle, Rankine cycle, cooling system, and synchronous generators connected to the grid. A hydraulic model of the seawater intake is developed to link cooling-water mass flow variations with satellite-derived sargassum occurrence data, to determine how these disruptions propagate through the plant and affect electrical output. The results demonstrate the model’s capability to link environmental disturbances and threats to measurable thermodynamic and power-system variables, providing insights into their propagation through and their effects on plant performance, offering a basis for assessing plant vulnerability and supporting the development of more resilient power-generation systems under environmental hazards.

Keywords: Marine law and policy, Law of the Sea and UNCLOS, International issues
Abstract: The legal landscape for oceans and climate has changed dramatically in recent years with the issuance of several international court advisory opinions, myriad U.S. executive orders, and geopolitical considerations. This paper presents how these intersect with the United Nations Convention on the Law of the Sea related to marine pollutants as emissions, state sovereignty and security, and deep seabed mining. The issuance of U.S. executive orders in 2025 and 2026, especially related to energy and offshore critical minerals, introduce additional layers of geopolitical considerations and the application of customary international law. This paper evaluates the international court findings through the lens of 2025 – 2026 U.S. executive orders and emerging pressures to open areas to deep seabed mining. A clearer understanding of how the law is evolving amidst these changes and increasingly complex geopolitics is important to inform ocean management, research, and security.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Data fusion, uncertainty quantification, and predictive modeling
Abstract: Autonomous exploration in stochastic marine environments, such as hydrothermal vent localization and plume tracing, requires robust planning under significant uncertainty. While Monte Carlo Tree Search (MCTS) has emerged as a standard simulation-based framework for informative path planning, its efficacy is fundamentally constrained by the selection of the planning horizon. Static or heuristic-based horizon scheduling frequently leads to structural instabilities, characterized by oscillatory trajectories and high stochastic variance. These limitations often result in premature exploitation or inefficient search depths that fail to align with the dynamic information density of the environment, ultimately degrading the reliability of autonomous underwater operations.

This research proposes a continuous adaptive framework that dynamically modulates the MCTS search depth by integrating reward-based confidence and Gaussian Process (GP) uncertainty. The methodology introduces a dynamically scaled spatial regularization term to encourage proximity to high-reward regions during exploitation while preserving sufficient search depth during early-stage exploration, stabilized by a confidence-driven phase mechanism with temporal smoothing. The system is evaluated in a simulated source-seeking scenario where an autonomous agent navigates a spatial field modeled by a GP. Performance is quantified using Average Simple Regret, measuring the expected differential between the true global maximum and the agent’s best-found observation.

Experimental results demonstrate that the belief-driven adaptive framework significantly outperforms exponential baselines in both stability and convergence. The proposed method achieves a faster, more monotonic decline in simple regret after the initial exploration phase, suppressing performance fluctuations associated with late-stage drift. By maintaining bounded variance across trials, the framework proves resilient against misleading local optima, offering a scalable solution for the OCEANS community to enhance the reliability of autonomous sampling and environmental monitoring in complex underwater domains.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Low-cost, scalable oceanographic sensors and platforms, Data fusion, uncertainty quantification, and predictive modeling
Abstract: Safe navigation for Autonomous Surface Vehicles (ASVs) requires both accurate obstacle detection and actionable risk assessment. We propose WARD (Water-surface Awareness and Risk Detection), a vision-only, risk-aware framework that can output safety assessments for low-cost ASVs. WARD fuses SegFormer semantic segmentation with Depth Anything V2 monocular depth using a novel continuous collision-corridor risk model. The risk model reasons jointly about proximity, lateral offset, and semantic class to produce human-interpretable safety alerts. An extended SORT tracker ensures stable Time-to-Collision predictions by filtering monocular depth noise. Operating at 13.7 FPS on a consumer-grade GPU, WARD provides discrete safety warnings (SAFE, CAUTION, IMMEDIATE) for resource-constrained ASVs.

Keywords: Autonomous underwater vehicles, Low-cost, scalable oceanographic sensors and platforms, Vehicle design
Abstract: A Cerulean Omniscan 450 SS sidescan sonar was packaged into a payload module and integrated with a small, low-cost, custom AUV (MIT's SeaBeaver III) and off-the-shelf ASV (EchoBoat) to enable high-resolution imaging with low-cost hardware and a modular payload. Major tasks included mechanical mounting, development of 3D-printed parts, wet-side and dry-side cabling, electrical distribution, software integration, sonar data logging, and sonar control on a Raspberry Pi 4 and Blue Robotics Navigator board. Bench tests, pool trials, and sea tests were conducted to evaluate power sufficiency, signal logging capability, mission sequence logic, hardware integrity, and software stability. We present an integrated payload system that is capable of collecting georeferenced sidescan sonar data from either a SeaBeaver AUV or EchoBoat ASV. These platforms establish a baseline capability for future work, including onboard machine learning algorithm development and improved localization framework in the absence of GNSS input.

Keywords: Autonomous underwater vehicles, Offshore structures, Vehicle design
Abstract: Underwater construction is difficult to automate since manipulation and localization are unreliable in marine environments, and assembled structures must withstand external disturbances such as waves and currents. We introduce URCHIN, a modular voxel system for autonomous underwater assembly using general-purpose robotic manipulators. Each voxel integrates grasp-optimized geometry, embedded neodymium magnets for passive self-alignment, and ArUco markers for visual localization. To characterize magnetic behavior, we propose a reciprocal cubic model that estimates pull force from separation distance, enabling precise control and motion planning. Experimental validation with a 10.16 cm (4 in) prototype demonstrates reliable snapping under positional offsets up to 4.44 cm (1.75 in) and confirms structural integrity through underwater bending tests. Autonomous trials achieve consistent snapping in 1D, 2D, and 3D configurations without precision placement, supporting the assembly of complex geometries. URCHIN provides a basis for scalable, reconfigurable underwater construction in dynamic marine environments.

Keywords: Sensor synergy, Maritime domain awareness and autonomous monitoring, Vehicle design
Abstract: The inspection of underwater infrastructure is not a trivial task, making it both time-consuming and expensive. These aspects limit the achievable inspection frequency, leaving room for harmful conditions to appear and worsen. One approach to reducing costs and inspection time is to use autonomous underwater vehicles, in which the selection of appropriate sensors is a significant step in the design process. Despite the sensors defining which inspection tasks the autonomous system can conduct, the process of selecting them is rarely described. There is therefore a need for structured procedures that enable reflection on how different sensors can assist in solving a given task and explain why certain sensor choices were made. This paper presents a procedure, inspired by Fault Tree Analysis, called Modality Tree Analysis (MT-A), which uses both logic and custom gates to describe and explore the relationship between task-specific attributes and sensor modalities in a tree-like structure. The procedure is demonstrated on a constrained inspection task focusing on subsea cables, pipelines, and offshore monopile foundations.

Keywords: Vehicle design, Vehicle performance, Remotely operated vehicles
Abstract: Unmanned surface vehicles (USVs) integrated with remotely operated vehicles (ROVs) have emerged as an effective solution for marine operations, including environmental monitoring, subsea inspection, and offshore maintenance. In such systems, the USV provides a long-endurance platform for navigation, communication, and coordination, while the ROV enables close-range sensing and intervention. This study presents the design, integration, and validation of a USV–ROV cooperative system for remote battery swapping of a seafloor observatory developed by the National Museum of Marine Biology & Aquarium (NMMBA), Taiwan. The observatory is scheduled for deployment approximately 5 km offshore at a depth of 70 m and is powered by rechargeable lithium battery modules with dual wireless charging interfaces. Conventional diver-based battery replacement introduces operational risks and limits maintenance frequency. The proposed system enables remote battery replacement and data retrieval while maintaining the observatory in situ. The system integrates a USV with navigation sensors and long-range communication, an ROV equipped with a custom battery swapping payload, a hydraulically actuated lifting platform, a winch system, a guided docking structure, and a shore-based control station. A key contribution is a compact ROV-mounted battery swapping mechanism that eliminates the need for complex manipulators. The mechanism employs a dual-gripper swing-arm driven by a linear actuator, enabling simultaneous insertion and extraction of battery modules in a single cycle. Its placement near the center of gravity reduces attitude disturbance and improves stability. Another contribution is a robust ROV launch and recovery system based on a hydraulic parallel-link platform, enabling reliable deployment and retrieval without human intervention. System performance was validated through integrated testing, including indoor tank experiments and outdoor aquaculture pond trials. The ROV successfully performed navigation, docking, battery exchange, and recovery under remote control, with a swapping time of approximately 1 minute. The results demonstrate reliable and repeatable operation without diver assistance.

Keywords: Open-source hardware, software, and community innovation, Rapid-deployment systems and modular instrumentation, Vehicle design
Abstract: Recent developments in the commercialization of marine robotics has resulted in lowering the costs of unmanned surface vehicles (USVs). This has promoted a wide spread adoption of vehicles like the BlueBoat from Blue Robotics for use in scientific data collection [1] and river monitoring [2]. The BlueBoat is a capable USV at a price that is an order of magnitude cheaper than previous small USVs such as the discontinued Clearpath Heron. At a cost of 5,000 dollars USD (2026) it is now feasible to deploy large fleets of these vehicles. This alone makes the BlueBoat an ideal candidate for building large scale autonomous fleets but its payload capacity as well as long mission duration capabilities further improve its position. This paper describes key modifications to the standard BlueBoat implemented by the MIT Marine Autonomy Lab to better leverage the effectiveness of these vehicles in large scale swarm operations and demonstrates early results attained through the operation of these vehicles showing promise for future research.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Rapid-deployment systems and modular instrumentation, Remotely operated vehicles
Abstract: Ocean-going autonomous systems play a crucial role in maritime missions, enhancing safety and efficiency through advanced sensors. The current state-of-the-art autonomous underwater vehicles (AUVs), remotely operated vehicles (ROVs), and unmanned surface vehicles (USVs), augmented by advances in sensor technology, offer dramatic improvements in cost and efficiency to support sensing operations in mine countermeasure missions. However, littoral environments such as estuaries and river deltas pose navigation challenges due to complex terrain and strong currents. The complex geographical features and rapidly changing environment, such as land barriers, icebergs, and tidal currents, may hinder their movement. For this reason, the Systems and Imaging Laboratory (SAIL) at the Harbor Branch Oceanographic Institute (HBOI), Florida Atlantic University, is developing a novel sensing approach to address these limitations through a new type of platform: Combining Robotics in Aerial and Nautical Environments (CRANE). CRANE is a multi-domain robotic system that utilizes waterproof unmanned autonomous vehicles (UAVs) with deployable underwater sensors. CRANE adopts a hybrid architecture that pairs a waterproof aerial drone with an underwater payload, connected via an optical fiber tether managed by a winch mounted on the drone. The payloads are retracted during aerial transit. To conduct sensing operations, the drone lands on the water surface and then idles its motors. The onboard winch then deploys the payload to the target depth to perform the intended tasks. The drone is waterproof and positively buoyant, enabling safe water landings and extended surface operations. CRANE has unique features to complement conventional ocean-going robotic platforms. - High energy efficiency: During sensing operations, the drone is idle on the water surface with the motors turned off. The winch and payload consume much less power. Therefore, the actual mission duration can be significantly longer than the aerial flight time. - Persistent link: With the drone resting on the water surface, the optical fiber between the underwater payload and the topside platform enables a continuous data link to a surface vessel or onshore control center.

Keywords: Vehicle performance, Low-cost, scalable oceanographic sensors and platforms, Hydrography / Seafloor mapping / geodesy
Abstract: Shallow coral reefs occupy a monitoring gap: too shallow for conventional vessel-based acoustic survey and too structurally heterogeneous for optical methods alone. Optical Structure from Motion (SfM) photogrammetry resolves coral morphology at millimeter to centimeter scale but degrades over texture-poor seagrass and sand, while multibeam echo sounding (MBES) provides substrate- and clarity-independent coverage at a resolution limited by beam footprint. We present a low-cost autonomous surface vehicle carrying dual parallel nadir cameras and a 240 kHz multibeam echo sounder, together with a workflow for co-registering the two surfaces over coral reefs in less than five meters of water. The multibeam bathymetry point cloud provides the georeferenced frame and metric scale, at standalone GNSS accuracy; the optical surface is registered into that frame, and seafloor targets of known dimension serve as an independent check on scale. Because both surfaces are placed in a common frame without independent geodetic control, inter-modality agreement is treated as shape agreement rather than absolute accuracy. Across triplicate surveys, the scale-invariant structural metrics repeated within 0.8 to 3.0 percent (coefficient of variation) in both modalities, with Arc-Chord Rugosity being the most repeatable. Structual relief metrics carried in physical units were more variable, consistent with their construction and acquisition method. Compared over the same reef, the two surfaces diverge systematically in structural magnitude: the acoustic surface reports higher mean slope and rugosity, and the optical surface results in higher maximum slope and total relief, a pattern consistent with the acoustic beam footprint. An optical-priority, acoustic-infill fusion approach is assessed for feasibility as a basis for future work.

Keywords: Naval architecture, Hydrodynamics
Abstract: Using 3D-printed models to advance hull form education and research Jackson Clendenin, Annslee Maloy, Corbin Freeland, and J. Travis Hunsucker. Ocean Engineering and Marine Sciences, Florida Institute of Technology, 150 W University Blvd, Melbourne, FL 32901. Traditional ship model construction and hull-form testing methods require advanced facilities with towing tanks, seakeeping basins, and machine shops. This infrastructure is needed to build and test final design models, however, the associated costs can be prohibitive for rapid prototyping in education, research, and design. To overcome these barriers, Florida Tech is developing methodologies to build self-propelled 3D-printed models and conduct low-cost open-water testing. The primary goal of these efforts is to provide a platform for education and student research in ship hull form design with lower costs, accessible facilities, and higher mobility. Initial designs are based on the hull DTMB 5514. Model hulls are 3D-printed and assembled in modules. The outer hull skin is finished to create a water-tight model. Control and data collection systems are modular and can be transferred between models. Testing occurs in the open water of the Indian River Lagoon adjacent to Florida Tech’s Mertens Marine Center. The models are controlled remotely by an onshore operator or a chase boat. Seakeeping and maneuvering tests are used to experimentally determine ship dynamics. Roll decay, inclining, and ship-wave interaction tests derive metacentric height and RAOs. Turning circle and spiral maneuvering determine turning stability and performance. These results will be validated with the publicly available benchmark data from the DTMB 5514 model tested at NSWC Carderock. These methods will allow organizations without specialized facilities to pursue hull-form research, design, and testing. This provides students the opportunity to test hull designs and can be expanded into industry for experimental validation during early design stages.

Keywords: Pollution monitoring, Hydrodynamics, AI/ML for oceanographic and environmental data analysis
Abstract: Rapid localization of pollutant releases in coastal waters is complicated by time-varying currents, wind forcing, shoreline geometry, and mixing, which continuously deform the measured plume. We investigate a model-free, single-agent reinforcement-learning approach to marine source seeking in time-varying concentration fields generated with the MIKE 21 hydrodynamic model for the coastal area of Cecina, Italy. The task is formulated as a partially observable control problem and addressed by a centralized Maskable Proximal Policy Optimization (PPO) policy using concentration and displacement histories, wind and local-current information, and plume-memory features. Spatial sensing is varied among central-agent-only measurements and ideal co-moving formations with one or two sensing rings. The three policies are compared with a field-climbing baseline across vehicle speeds, wind-forcing conditions, and plume ages.

Keywords: Under-ice robotics, sensing, and operations, Remotely operated vehicles, Autonomous underwater vehicles
Abstract: Autonomous Underwater Vehicles (AUV) and Remotely Operated Vehicles (ROV) are extremely valuable and often essential to expeditionary science, especially in harsh and extreme environments such as the deep ocean and polar regions. However, field robots face several challenges in these environments, including physical and power limitations, sensor degradation and noise, and operating under unforeseen, unstructured conditions. These factors can significantly hinder or even prevent successful missions. In this extended abstract, we present a field experience report of near-shore benthic surveys conducted under landfast ice using MOLA, a portable 6-DOF hybrid ROV/AUV. We discuss the system configuration, perception and navigation modules, and deployment details during remote operations. We further highlight the challenges encountered, lessons learned, and datasets collected in the surveys.

Keywords: Vehicle navigation, Autonomous underwater vehicles, Real-time decision systems for ocean robotics and autonomous operations
Abstract: Self-localization is the most important and indispensable technology for AUVs. While underwater high-precision self-localization is a challenge, acoustic positioning, currently the dominant method, suffers from numerous problems such as multiple reflections, thus creating a demand for alternative methods. This research aims to establish two methods of position estimation for autonomous underwater vehicles (AUVs) not using acoustics that are “Motion Estimation from Image (MEfI)” , which is an advanced “Structure from Motion(SfM)” method, and “Relative self-position Estimation from Image(REfI)”, which uses image maps and processes image feature values using “Artificial Intelligence(AI)”. Using image-based self-positioning technology, the following series of sequence control tests were successfully conducted in a real sea area: 1. Docking state with the underwater charging station (communication with the underwater charging station) 2. Detachment from the underwater charging station 3. Navigation (survey) [Self-position estimation using only images] 4. Docking to the underwater charging station using position estimation using only images 5. Communication between the AUV and the underwater charging station and charging of the AUV This series of sequence control tests was performed using only self-position estimation using images, without acoustic self-position estimation.

In this sequence of control operations, the AUV charges, retrieves data, and communicates underwater at an underwater charging station. It was demonstrated that long-term surveys can be conducted without lifting the AUV from the water into the air, and using image-based self-localization, which does not require sound.

Keywords: Vehicle navigation, Sound propagation and scattering, Autonomous underwater vehicles
Abstract: This work studies how refraction-induced bias on acoustic ranging affects multi-agent collaborative localization in a range of oceanographic conditions and spatial scales. While multi-agent range-aided navigation, which uses range measurements to either fixed infrastructure or other agents, is a promising solution to the challenges of large-scale underwater localization, its accuracy depends strongly on the quality of range measurements. Sound speed variability induces refraction (bending) of acoustic rays, yet, for algorithmic tractability, standard sensor fusion pipelines assume straight-line propagation. This refraction systematically biases range measurements to be longer than the straight-line assumption predicts. However, the effects of this bias on multi-agent collaborative localization on kilometer scales remains unexplored. We present a series of simulated experiments with several agents operating over kilometer scales. The simulation uses HYCOM reanalysis data to recreate realistic oceanographic conditions, ray tracing to generate refraction-informed ranges, and a centralized multi-agent factor graph estimator to quantify the resulting measurement bias on estimated trajectories. Preliminary results indicate that refraction-induced bias can induce significant degradation of estimated trajectories, particularly in regions with sharp sound-speed gradients. We also share the simulation environment to support further studies https://github.com/UMich-RobotExploration/manta-ray.

Keywords: Autonomous underwater vehicles, Vehicle navigation, Vehicle performance
Abstract: High-precision, servo actuators are indispensable for subsea operations that demand accurate positioning, responsiveness, and continuous repeatability under extreme subsea conditions. This abstract outlines the fundamental principles and advancements of a readily adaptable 6,000-meter ocean rated, pressure tolerant rotary actuator known as the HD Servo. The HD Servo utilizes robust corrosion resistant material and integrates a high-torque brushless motor, zero backlash gear train, and high-resolution feedback encoder that are contained in a compact, pressure-compensated housing. The attached integrated closed-loop controller is located in an environmentally sealed pressure tolerant enclosure that continuously monitors actuator feedback and dynamically adjusts motor drive signals for quick movements and maintained positions, enabling both precise angular motion and robust disturbance rejection in a compact package with a single external connection. Assembled HD Servo, bench test fixture, pressure test, and AUV Sentry installation The HD Servo is primarily adapted for use on Autonomous Underwater Vehicle (AUV) Sentry prioritizing long-term reliability, field serviceability, and performance at high pressure, low temperature environments. Bench testing and field trials with the HD Servo demonstrate the actuator’s performance when operating the AUV’s control surfaces at high frequencies during terrain following missions. High resolution bathymetric and side scan sonar data collected in challenging and high relief terrain demand a high degree of responsiveness and accuracy in order to maintain attitude control for 24-hour missions. The integrated controller executes custom motion profiles and provides health monitoring diagnostics in real time and inductive absolute rotary encoder provides rapid and accurate positions that resist electrical noise and requires no preliminary calibrations or movement on startup. Placing the sensitive control electronics in a 1-atmosphere environmentally sealed enclosure and passing control and power signals through a pressure boundary PCB through the two separate volumes dramatically improves the reliability, serviceability, and compact nature of the HD Servo. Th

Keywords: Remotely operated vehicles, Dynamic positioning, Autonomous underwater vehicles
Abstract: Precision positioning remains a critical operation in autonomous underwater manipulation, particularly when operating Remotely Operated Vehicles (ROVs) under strict actuator constraints. This paper proposes a cascaded control architecture that integrates a kinematic model predictive control (MPC) approach with a prioritized, higher-frequency velocity controller. While dynamic MPC leverages the full system model, it is often difficult to model uncertainty in hydrodynamic parameters. To address this, we use a sliding virtual target approach that shapes the cost function vector field towards a smooth spiral rather than a straight-line path. This predictive layer is coupled with an internal discrete PID execution layer that employs a priority-based thrust allocation logic, ensuring that vehicle stability (roll/pitch) is preserved even when thrusters reach physical saturation limits. Simulation results obtained with the marine simulator Stonefish demonstrate that this architecture can achieve centimeter-level positioning accuracy and stable heading alignment for manipulator tasks.

Keywords: Vehicle design, Vehicle performance, Dynamic positioning
Abstract: Research vessels handling heavy scientific equipment on confined decks face challenges associated with limited deck space, vessel motion, and the need for flexible equipment handling. This paper presents the design, fabrication, and onboard evaluation of an omnidirectional auxiliary vehicle based on Mecanum wheels. The full-scale platform was designed for a 3,000-kg payload and provides holonomic motion and zero-radius turning for maneuvering in confined deck areas. During deck trials aboard R/V Legend, the vehicle carried 750 kg of ballast, and the force-gauge reading reached its 300-kgf (approximately 2.94-kN) measurement limit in both the longitudinal and lateral pulling directions. Because the gauge saturated, the maximum drawbar pull was not quantified beyond this value. The vehicle was also integrated with a gravity-corer support fixture and evaluated in a shipboard handling procedure that avoided the installation and removal of conventional deck sliding rails.

To investigate future supervised autonomous assistance, a vision-based position-keeping method was implemented on a downscaled Mecanum-wheel prototype. A downward-facing OpenMV Cam H7 Plus uses phase-correlation-based image displacement estimation to determine deck-relative translation without GNSS or external positioning infrastructure, while a discrete PID controller generates corrective wheel commands. Repeated indoor disturbance tests showed qualitatively consistent corrective translational behavior, although residual yaw drift remained because heading was not estimated or controlled. These results demonstrate the feasibility of combining omnidirectional heavy-load transport with future deck-relative position-keeping assistance, while full-scale integration, yaw correction, and formal operating limits remain subjects for further development.

Keywords: Ocean energy, Automatic control, Offshore structures
Abstract: This paper presents a real-time model predictive control (MPC) scheme for a floating wind-wave combined system (WWCS). Every 0.1 s, the scheme updates one power-take-off (PTO) damping coefficient for each of three wave energy converters (WECs). The objective is to reduce platform pitch and roll motions. Loads from a coupled AQWA-OpenFAST simulation are separated into PTO and non-PTO moments. The non-PTO moments are predicted over a short horizon with an autoregressive (AR) model updated by recursive least squares (RLS). A constrained quadratic program is then solved with the HPIPM interior-point solver. The fixed-damping baseline uses B_fix = 6 x 10^5 N s/m for all three WECs. Relative to this baseline, the MPC scheme reduces pitch and roll standard deviations by 35.2% and 43.5%, respectively.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Autonomous underwater vehicles, Automatic control
Abstract: Dexterous manipulation with small-scale underwater robots is essential for tasks such as environmental sampling and infrastructure inspection, but requires accurate control of the forces exerted by the robot. Dedicated submersible force/torque sensors are often too expensive and bulky for compact platforms. Building on our prior work on model-based wrench estimation using only onboard motion sensor measurements, we present the first implementation of an admittance control pipeline on a small-scale underwater vehicle-manipulator system without using a dedicated force/torque sensor. The estimated interaction wrench serves as force feedback for an admittance controller that regulates the interaction force. In a systematic evaluation this pipeline is compared to a baseline pipeline, which uses a dedicated force/torque sensor. We demonstrate the feasibility of this approach in a human-robot collaboration task, which highlights its relevance for real-world applications.

Keywords: Foundation models and large-scale training for marine applications, Automatic control, Maritime decarbonization and eco-efficient operations
Abstract: This paper models a single winter crossing of the North Sea, from the Port of Tyne to Kristiansand, as a minimum-fuel routing problem on a static H3 hexagonal ocean graph carrying hourly time-varying edge costs, and solves it with a masked double deep Q-network. Every edge is priced in main-engine fuel consumption by a Holtrop and Mennen calm-water model, a Fujiwara wind term and a STAWAVE-1 added-wave term, all driven by ERA5 and Copernicus Marine fields. Two variants are trained under identical conditions, a bare hull and the same hull carrying two rigid wingsails. Because the learning agent and a time-expanded Dijkstra planner are charged from the same edge-cost table, the optimum of the discretised problem is known exactly rather than estimated. The best policy in each variant reproduces that optimum, and the ten-seed mean lies within 0.20% of it with wingsails and 0.13% without. The trained agent saves 8.2% of fuel against the distance-only haversine route, while the wingsails save 4.9% on the fuel-optimal route but only 3.9% on the distance-only route, so route quality and wind assistance compound.

Keywords: Automatic control, Real-time decision systems for ocean robotics and autonomous operations, Vehicle navigation
Abstract: : 본 논문은 해양 사물인터넷(IoT) 임무를 위한 무인 수상 우주선(USV)의 추진 이상 탐지 및 협력 제어를 위한 통합 프레임워크를 제시합니다. 비정상적인 추진 데이터는 해상에서 얻기 어렵고 비용이 많이 들기 때문에, 장기 단기 기억 오토인코더(LSTM-AE)는 정상적인 추진기 진동 데이터만을 사용해 훈련하며 재구성 오류로 인한 이상 징후를 감지합니다. 진동 데이터는 1 kHz에서 3m 실험용 USV에서 수집되었습니다. 정밀도-회상 분석은 이상도 임계값을 5.31로 선택하여 정밀도 0.803, 회상 능력 0.795, AUROC 0.916을 얻었습니다. 이상 현상 출력은 2자유도 USV 시뮬레이터, ANN-PID 방향 제어기, 반발력, 방향 및 인력 구역을 이용한 규칙 기반 군집 행동과 통합되었습니다. 두 척의 USV의 정상적인 해상 시험 궤적을 사용해 시뮬레이션을 &#

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Vehicle design, Autonomous underwater vehicles
Abstract: Hydrodynamic coefficients obtained from towing-tank tests or computational fluid dynamics (CFD) rarely match the as-built vehicle, which motivates recalibration from routine sea-trial navigation logs of an autonomous underwater vehicle (AUV). Physics-informed neural networks (PINNs) are in principle well suited to this inverse problem, but their accuracy on operational logs remains inconsistent. This study develops a six-degree-of-freedom PINN for a torpedo-shaped AUV and finds that estimation accuracy is primarily limited by the information retained after differentiation of the navigation measurements, rather than by the network architecture or the optimization procedure. Given exact accelerations, the network recovers the coefficient set to 0.42 % mean absolute error. Given measured accelerations, the error increases to 33.45 %. A per-equation derivative signal-to-noise ratio is therefore defined. The criterion is computed from the log alone and requires neither ground truth nor a sensor datasheet, and it is used to set the state-network bandwidth, to weight the residual and to screen the coefficient set. Across 135 minutes of sea trials it identifies before training which coefficients are recoverable, and the best of four fits reduces the open-loop position error of a straight run from 21.9 m to 8.8 m.

Keywords: Array signal processing and array design, Sonar signal processing, Signal coherence and fluctuation
Abstract: In shallow-water active sonar systems, underwater moving target detection is severely hindered by strong residual reverberation.Platform motion introduces azimuth-dependent Doppler shifts to seabed scattering returns, causing spectral broadening that overlaps with target echoes and cannot be resolved by temporal processing or spatial beamforming alone. This paper proposes a space-time adaptive processing (STAP) method utilizing a continuous waveform (CW) to jointly exploit spatial and temporal degrees of freedom for reverberation suppression and moving target detection.

A temporal reverberation field model for a moving monostatic sonar in a range-independent ocean acoustic waveguide is established. Analytical expressions for the signal-to-reverberation ratio (SRR) gain are derived by decomposing the gain into the output-to-input target echo ratio (OITR) and the output-to-input reverberation ratio (OIRR). This framework reveals two fundamental advantages of CW. First, CW's high Doppler sensitivity produces a narrow temporal matched-filter mainlobe, yielding a significantly higher OIRR for moving targets than Doppler-tolerant wideband waveforms. Second, in dispersive waveguides, modal group-velocity differences cause temporal smearing of wideband echoes, degrading the OITR for LFM and MLS waveforms, whereas CW echoes remain immune to modal dispersion, preserving peak intensity.

Simulation results in a 30-m shallow-water waveguide at 1 kHz confirm this analysis. After STAP, CW achieves an output SRR exceeding 35 dB for target velocities above 1.5 m/s, surpassing both LFM and MLS at 400 Hz bandwidth. While MLS offers precise range-velocity estimation, its high uniform sidelobes introduce range ambiguity that may obscure weak echoes.

Although CW inherently lacks fine range resolution, this limitation can be mitigated by exploiting STAP-derived Doppler and direction-of-arrival estimates within a Kalman filter framework for continuous target trajectory tracking. Future work will implement this tracking approach and validate the proposed method with shallow-water sea-trial data.

Keywords: Acoustic telemetry and communication, Signal coherence and fluctuation
Abstract: The JANUS underwater acoustic standard (STANAG~4748) employs a legacy rate one-half convolutional code whose coding gain is limited at the minimum packet size of 64 information bits and 144 coded bits. This paper presents an equal-condition comparison of five forward error correction (FEC) codes---convolutional, turbo, low-density parity-check (LDPC), and two cyclic redundancy check (CRC)-aided polar variants---using actual JANUS frequency-hopped binary frequency-shift keying (FH-BFSK) waveforms with noncoherent energy detection over additive white Gaussian noise and frequency-selective channels. Unlike prior baseband-equivalent studies, this approach preserves the actual FH-BFSK waveform, hopping pattern, and chip timing, and therefore captures the noncoherent detection loss inherent in JANUS. Noncoherent detection compresses coding-gain differences: at a one-percent block error rate (BLER), turbo and polar achieve only 0.2 to 0.3dB coding gain over the convolutional code, while LDPC remains 1.2dB inferior. Turbo exhibits a marginally steeper waterfall at high signal-to-noise ratio, while adaptive successive cancellation list (Adaptive-SCL) polar decoding reduces per-frame decoding cost to 58 percent of turbo's fixed cost via CRC-aided early termination. Under a three-tap frequency-selective Rician channel the compression becomes complete: at ten-percent BLER, turbo and polar fall within 0.06dB of the convolutional code, and all three share an error floor near 4.4 percent caused by residual inter-chip interference, while LDPC falls 2.7dB behind. FEC selection for JANUS is therefore governed by decoding complexity rather than block error rate; further gains require receiver-side equalization, not stronger codes.

Keywords: Acoustic telemetry and communication, Sound propagation and scattering
Abstract: In underwater acoustic OFDM systems, each multipath component may undergo a different Doppler scaling, and compensating only the dominant path introduces residual intercarrier interference on the remaining paths. This paper proposes Superimposed HFM Peak Association (SHPA), which pairs peaks detected in the up-chirp and down-chirp correlation outputs so that the propagation delay cancels and the Doppler factor of each path is obtained directly. No additional FFT is required beyond the two correlations already performed in any superimposedHFM receiver; only a low-cost peak-association step, quadratic in the number of paths, is added. On a 19-ray channel derived from Bellhop ray tracing with a measured sound speed profile, a multiresampling MMSE receiver driven by SHPA approaches the ideal bound within 1 dB and provides significant gain over singleresampling. Cross-validation on an additional measured-profile channel and a robustness analysis over Doppler spread, path spacing, and detection threshold validate the proposed estimator

Keywords: Sonar signal processing, Model-based signal processing techniques, Sound propagation and scattering
Abstract: Sonar systems routinely operate in unique and challenging environments that demand novel waveform designs to optimize system performance. In practice, sonar systems rely on transmitting spectrally flat waveforms with constant moduli, such as Linear Frequency Modulated (LFM) waveforms, to detect targets in complex undersea environments. An LFM waveform maintains a uniform energy distribution across the transmit bandwidth; regardless of noise and clutter Power Spectral Density (PSD) fluctuations. This misallocation of spectral energy wastes transmit power in frequency bands with high noise and clutter thus reducing the Signal to Noise Ratio (SNR) of a received echo. In contrast, Matched Ensonified (ME) waveforms leverage the mutual information between transmit signals and channel-target responses to design waveform spectra that maximize target detection probability. The magnitude squared spectra of ME waveforms, also known as the Energy Spectral Density (ESD), routinely surpasses the detection performance of LFM waveforms. While ME waveforms promise superior detection performance, the mathematical process that defines their spectral shape does not guarantee the waveform will possess a constant modulus. Practical sonar systems cannot readily accommodate these amplitude fluctuations without significant input power backoff to preserve transducer health. Consequently, theoretical detection gains are offset by the reduced transmit power needed to maintain hardware integrity. This work utilizes the Multi-Tone Sinusoidal Frequency Modulated (MTSFM) waveform model to approximate the ESD of ME spectra to balance detection performance with transducer amenability. For point target simulations, MTSFM waveforms that approximate ME ESDs exhibit superior detection capabilities over LFMs. We build upon previous analysis by modeling the backscattered target strength of several complex elastic targets in environments with variable noise and clutter to evaluate the detection capabilities of MTSFM waveforms for realistic detection scenarios. By synthesizing physics-based models with adaptive waveforms, we validate the efficacy of MTSFM waveforms that approximate ME ESDs across a diverse range of realistic target scenarios.

Keywords: Passive observing sensors, Sonar signal processing, Distributed hydrophone arrays and networked acoustic systems
Abstract: This paper investigates a passive sonar localization framework that potentially integrates asynchronously distributed horizontal and vertical hydrophone arrays (HHA and VHA) in shallow-water environments. Passive localization in these regions is inherently challenging due to complex multipath propagation and the technical difficulty of maintaining precise time synchronization across distributed sensor nodes. The proposed methodology explores the synergy between warped frequency spectral alignment for HHA, which addresses modal dispersion, and multipath-based striation analysis for VHA, which leverages waveguide invariant properties. By analyzing the theoretical performance and conducting feasibility studies under various sound velocity profiles and seabed conditions, this research evaluates the potential of an integrated, "sensor-cooperation-free" approach. This asynchronous configuration aims to overcome the inherent limitations of individual array systems—such as range-depth ambiguity—without the hardware overhead of a common-clock system. The findings provide critical insights into how such a decentralized architecture can enhance the reliability and scalability of acoustic monitoring in complex maritime environments. This study offers a practical foundation for future autonomous underwater vehicle (AUV) networks and distributed maritime surveillance systems.

Keywords: Passive and active acoustic monitoring, Distributed hydrophone arrays and networked acoustic systems, Acoustical oceanography
Abstract: Passive acoustic measurements collected on the Washington continental shelf during June–August 2025 are used to examine the influence of subsurface sound channels on the ambient sound field. Observations were obtained using dual vertical line array (VLA) systems deployed at two sites, each consisting of a compact 4-element subsurface array and a near-bottom array with 30 cm spacing. These collocated arrays sample ambient sound at near-surface and near-bottom depths, capturing depth-dependent variability associated with subsurface ducting in a dynamic coastal environment. Prior work in this region established coastal upwelling produces persistent subsurface duct. The spectral analysis reveals intermittent acoustic energy enhancements consistent with ducted propagation, varying on tidal (hours) to synoptic (days) timescales. Characterization of the acoustic angle of arrival (AoA) was predominantly near-horizontal angles across the array depths. The ambient noise field tracks local wind to allow passive wind speed retrieval. Results highlight the role of stratification in shaping coastal soundscapes and demonstrate the utility of compact passive VLAs within multi-platform observational programs.

Keywords: Data fusion, uncertainty quantification, and predictive modeling, Acoustical oceanography, Benchmark datasets, reproducibility, and open-source frameworks
Abstract: Interpolated sound speed fields built from Argo float profiles underpin naval acoustic prediction, autonomous vehicle mission planning and seafloor geodesy. What those products say about their own reliability varies widely. Some carry no error estimate at all, and others carry one that has never been checked against independent measurements. A practitioner asking for sound speed at a point may get a number with little sense of how far it might be from the truth.

This paper presents kernel-weighted spatiotemporal interpolation (KWSI), which returns three things at every point queried: a sound speed estimate, an error estimate, and a measure of observational support, meaning how much float data actually stood behind that value. The error estimate is built from four named sources, three describing what happens within a single float profile and one covering what is lost by estimating at a place no float visited.

The error estimate was checked against float data held back from the fit, across 20,779 withheld cycles over the Bay of Bengal from 2011 to 2020, and it holds. Breaking it into its parts gives the central result. Sensor precision accounts for 0.0003 percent of the total, while the term covering what is lost between floats accounts for just under 99 percent. The reliability of one of these fields is set by how densely the ocean is sampled, not by how good the instruments are.

That dominant term takes the same value everywhere at a given depth, so the reported error looks nearly flat across the map while the amount of data behind it varies by more than tenfold. An error bar alone therefore says nothing about how well observed a value is, which is why support is reported as a field of its own. Code is available at github.com/Calvinxc1/argo-data-interpolation.

Keywords: AI/ML for oceanographic and environmental data analysis, Signal coherence and fluctuation, Acoustical oceanography
Abstract: NOAA Fisheries initiated a strategic initiative to transition active acoustic data processing and analysis to the cloud, where we are developing and implementing automated data pipelines, and machine learning and artificial intelligence methods to revolutionize echo classification. Species-level echo classification of highly diverse fish assemblages is currently constrained by the availability of labeled, biologically ground-truthed data. Such data can be particularly challenging to collect for high-diversity fish assemblages at untrawlable habitats, such as reefs. Camera-based optical surveys are commonly used in fishery-independent sampling of reef fishes, but are limited by water visibility, the ability to enumerate fish in dense (and often mixed species) schools, and encounter rates for fishes associated with relatively small, isolated habitat patches. A pilot survey was conducted in the southeastern U.S. utilizing broadband active acoustics to augment the Gulf Fishery-Independent Survey of Habitat and Ecosystem Resources, a primary data source for stock assessments of reef-associated fishes in the region. We are developing a weakly-supervised modeling framework to biologically ground the classification of broadband spectra from reef fish targets by applying species richness from optical data towards hyperparameter tuning of hierarchical agglomerative clustering. The proposed modeling framework provides a tool for partitioning acoustically-derived abundances of highly-diverse fish assemblages with minimal requirements for ancillary biological data.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Array signal processing and array design, Oceanographic instrumentation and sensors
Abstract: This study presents experimental validation of a low-cost, 3D-printed sonar head, demonstrating the potential of additive manufacturing for rapid development of underwater acoustic systems. A 16-element baffled circular hydrophone array was fabricated using 3D printing, with emphasis on acoustic performance. Controlled measurements conducted in a hydroacoustic tank show that the proposed baffle structure achieves at least 20 dB attenuation in opposite directions, significantly enhancing array directivity over the 10–20 kHz frequency range. A sea trial in a shallow-water environment demonstrate the system’s capability for range detection of coastal features, including rocky boundaries. These results support the feasibility of a low-cost 3D print-based sonar system. The study highlights both the advantages and current limitations of using 3D printing in the development of a passive–active sonar unity.

Keywords: Sonar imaging, Hydrography / Seafloor mapping / geodesy, Autonomous underwater vehicles
Abstract: Sidescan sonar is a common sensor for both manned and autonomous marine exploration and mapping, yet very few methods build upon or exploit the geometric projection model of the sensor. As sidescan sonar is limited to a one-dimensional range measurement, many approximations are frequently used, including the long-standing assumption of a flat seafloor. Rather than make similar approximations, this paper focuses on a multi-view geometry based approach and formalizes a two-view geometric constraint and proves that a shared feature is constrained to a locus within the intersection of a sphere and a plane. In addition, we characterize what governs the size of the ambiguity locus through Monte Carlo simulation that is grounded in real aperture and mounting geometry for both a surface vessel and underwater vehicle. We translate additional simulations of relative trajectories for both vehicle platforms into concrete survey-planning guidance. Our results show that locus length is strongly governed by elevation misalignment, and peaks at a moderate oblique crossing angle of approximately 20 degrees, with minimal locus lengths obtained at near parallel and anti-parallel passes.

Keywords: AI/ML for oceanographic and environmental data analysis, Imaging and vision, Foundation models and large-scale training for marine applications
Abstract: Vision-language models (VLMs) enable robotic perception by associating visual observations with natural-language concepts. Yet their performance in coastal environments remains largely unexplored. We introduce a densely labeled coastal dataset containing more than 1,000 images collected across seven missions in three regions of Oahu, Hawaii, with 18 semantic classes and over 7,400 annotated instances. We evaluate seven modern VLMs through three complementary experiments measuring text-to-mask, mask-to-mask, and mask-to-text alignment. Broad landscape classes are generally recognized more accurately than conventional object and coastal classes, with coastal concepts presenting the greatest challenge. However, comparisons of shared conventional classes across coastal and terrestrial datasets reveal no consistent performance difference attributable solely to environmental context. Mask-to-mask matching also remains similar across conventional and coastal classes, while alternative textual labels substantially improve recognition of several coastal concepts. These results suggest that lower performance on coastal classes (at least on the objects/query categories evaluated) is heavily influenced by segmentation and linguistic representation.

Keywords: Foundation models and large-scale training for marine applications, Imaging and vision
Abstract: Monitoring kelp forest health is critical for conservation, yet traditional methods require either specialized hardware or extensive manual annotations to train a supervised segmentation model. This study focuses on leveraging the zero-shot capabilities of the SAM3 model in the context of marine conservation, automating kelp canopy detection from RGB UAV imagery without the need for any domain-specific training. Using natural language prompting, we demonstrate that SAM3 is capable of achieving segmentation performance comparable to supervised machine learning methods. Because no pixel-level ground truth exists for the imagery used, we evaluate in two stages: agreement between methods across the full survey, and a hand-scored sample of images and points drawn from each outcome. Across a 476-image survey, SAM3 and a supervised CNN baseline agreed closely on canopy segmentation wherever both detected it, reaching an intersection over union of 0.81 on the 105 images where both detected canopy. At the image level, SAM3 reached a precision of 0.96, exceeding that of the supervised baseline. We further show that text prompting is sensitive to the vocabulary used, and that supplying annotated examples of the target appears to improve detection of small, camouflaged targets in a preliminary test. These findings suggest that foundation models can make ecological monitoring significantly more accessible.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Imaging and vision
Abstract: The ocean's biological carbon pump plays a critical role in mitigating global warming, sequestering approximately 10.2 Pg C yr⁻¹ through the sinking of phytoplankton-derived organic carbon, in the form of detrital particles. The efficiency of this pump varies across ocean regions, shaped by differences such as seasonal forcing and phytoplankton community composition. Eastern boundary upwelling systems are hotspots of productivity, and their role in the ocean carbon cycle have been extensively studied in temperate locations, such as the California Current. Their tropical counterparts, particularly the Gulf of Guinea, remain undersampled and underrepresented in global carbon cycle models, and we do not know whether the mechanisms of carbon export in these tropical ecosystems are shared with their temperate counterparts. A new, multi-year observational program in the Gulf of Guinea is now providing key data on the seasonal mechanisms of coastal carbon export. By comparing particle fluxes in the Gulf of Guinea to the California Current, we test whether variability in phytoplankton community structure and sinking particle composition are similar across these ecosystems and determine whether these dynamics diverge between tropical and temperate upwelling environments. To quantify particle fluxes and classify the ecological source of sinking particles, we deployed low-cost, open-source time-lapse imaging cameras (SnoCam+) on drifting sediment traps. SnoCam+ collects images of particles as they settle into a collection tube, resolving hourly variability in fluxes across the entire deployment period. In parallel, we characterized the surface phytoplankton communities using PlanktoScope, allowing us to directly link surface community composition to the composition and flux of sinking particles below. We trained CNN transformer models to automate the classification of these datasets. This early data offers an initial window into the size, abundance, and flux rates of sinking particles across contrasting upwelling ecosystems. Ultimately, the results of this study will deliver datasets and advances that will meaningfully improve the representation of tropical African waters in global biogeochemical models.

Keywords: Imaging and vision, Remote sensing and field campaigns in polar regions, Low-cost, scalable oceanographic sensors and platforms
Abstract: This work describes the design, integration, and field deployment of a dual-configuration real-time laser scanning system aboard MBARI's MiniROV, built for high-resolution 3D Arctic seafloor mapping. The system integrates two structured-light units: one downward-looking and one forward-looking; paired with SPLASH-SegFormer, a transformer-based segmentation pipeline that extracts laser stripe centerlines at 12Hz under challenging polar conditions. The system was deployed during MBARI's 2025 Arctic Expedition in the Canadian Beaufort Sea, where submarine permafrost morphology was surveyed across five pingos, under extreme turbidity and low ambient light. Seventeen surveys were completed successfully, producing real-time 3D reconstructions of both seafloor topography and vertical wall structures at millimetric resolution. These results show that learning-based laser scanning can operate reliably in real ocean conditions and point toward its broader use in monitoring climate-sensitive seafloor morphology.

Keywords: AI/ML for oceanographic and environmental data analysis, Pollution monitoring, Imaging and vision
Abstract: Unmanned aerial vehicles can survey more beach area than field teams can cover efficiently. Natural backgrounds and partly buried objects, however, complicate automated segmentation. We present the Pixel-level Litter Area Segmentor (PLAS-Net), a lightweight model that produces an individual mask for each item in high-resolution aerial imagery. PLAS-Net adds three feature-fusion and attention modules to a YOLO26n-seg baseline. A survey at Aow Luek Bay, Thailand, produced a 0.17-cm-per-pixel orthomosaic and 1,300 non-overlapping image tiles. On the 395-image evaluation split, PLAS-Net achieved 69.7 percent mask precision and 58.7 percent mean average precision at an intersection-over-union threshold of 0.5. Mean average precision averaged over thresholds from 0.5 to 0.95 was 46.8 percent. These three metrics exceeded the baseline by 16.5, 5.1, and 2.5 percentage points. The parameter count increased from 2.69 to 2.73 million. Scene-wide inference also produced size-frequency and footprint summaries from the predicted masks.

Keywords: Imaging and vision, Classification and pattern recognition (parametric and non-parametric), AI/ML for oceanographic and environmental data analysis
Abstract: Artificial intelligence (AI) and computer vision techniques have been widely adopted for fish detection, classification, and tracking. However, analyzing fish in natural underwater environments remains challenging due to low-light conditions, variable water clarity, frequent occlusions, visually similar species, and irregular fish movements. Reliable multiobject tracking is particularly difficult under these conditions. Non-linear motion and occlusion can lead to identity switches and fragmented trajectories. In this work, we use the Gulf Fishery-Independent Survey of Habitat and Ecosystem Resources Dataset 2024 (GFISHERD24), a large-scale underwater dataset containing more than 155 fish species. We develop a tracking-bydetection framework that combines a YOLOv26-based detector with an enhanced BoT-SORT tracker. To better account for nonlinear fish motion, an Unscented Kalman Filter is incorporated for motion estimation, while Re-Identification-based appearance features are used to improve identity association. The proposed framework is evaluated on 45 GFISHERD24 test videos under both single-class and multi-class settings. For single-class tracking, the proposed method achieves 59.9% MOTA and 50.843% HOTA, with an identity-switch rate of 1.30%. For multi-class tracking, it achieves 53.5% MOTA and 48.244% HOTA, with an identity-switch rate of 1.10%. These results demonstrate improved tracking robustness and identity preservation under challenging underwater conditions. The proposed framework supports automated species-level abundance estimation by enabling individual fish to be tracked consistently and counted only once.

Keywords: Imaging and vision, AI/ML for oceanographic and environmental data analysis, Benchmark datasets, reproducibility, and open-source frameworks
Abstract: Underwater object detection degrades when training data do not match deployment conditions - a frequent constraint for Autonomous and Remotely Operated Vehicle (AUV/ROV) operations, where collecting labelled imagery across the expected range of turbidity is expensive. Two approaches compete: restoring the image at inference, or synthesising turbidity during training. We compare them while keeping the detector and all of its training settings held identical, so that only the image transformation changes, training on clear-water images and evaluating on turbid ones.

On the Real-world Underwater Object Detection (RUOD) dataset, we first measure a strong association between which species appear in an image and how turbid that image is (Cram'er's V = 0.41): clear images hold mostly reef species and turbid images mostly seabed ones, so comparing accuracy between the two partly measures where the animals live. The association is weaker on datasets covering fewer habitats and species, at 0.11 on the Detecting Underwater Objects (DUO) dataset. Our physics-informed augmentation improves mean average precision (mAP@50) by 2.8 points, and by 4.1 with random erasing applied on top of it. Against enhance-then-detect methods, we find no significant difference from multi-scale fusion (p = 0.68), which costs 192 ms per frame on the tested CPU, while calibrated dark-channel restoration does not significantly improve over the no-augmentation reference. The augmentation acts only at training time and adds no computation at deployment.

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications, Benchmark datasets, reproducibility, and open-source frameworks
Abstract: Benthic ecologists are inundated with high-resolution imagery, making annotation a significant and costly bottleneck. While computer vision has shifted workflows toward automation, significant challenges still persist: a lack of interoperability between specialized ecological software, limitations in supported annotation geometries (e.g., points, polygons, masks), and restricted user control over the local training and deployment of deep learning models. The CoralNet-Toolbox aims to address these bottlenecks and provides a localized, end-to-end machine learning ecosystem. It handles diverse data ingestion and supports multiple raster types common to ecological analysis, facilitating computer vision tasks from classification to panoptic segmentation. The system integrates foundation-model-assisted annotation, uncertainty-ranked review, and embedding-based dataset-scale labeling. Ultimately, this architecture ensures that automation acts as a multiplier for expert throughput without displacing their authority, establishing verifiable provenance for AI-assisted ecological science.

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications, Benchmark datasets, reproducibility, and open-source frameworks
Abstract: This paper presents a comprehensive Sim-to-Real approach based on CycleGAN and multi-condition data augmentation to address the critical shortage of annotated underwater training data required for autonomous sea urchin detection. Sea desertification, primarily driven by excessive sea urchin overgrazing, severely degrades coastal seaweed beds. While Autonomous Underwater Vehicles (AUVs) offer a crucial solution for automation, removing the physical dangers and high costs of manual diver removal, training robust deep learning models like YOLO remains a significant bottleneck. This is due to the extreme difficulty of acquiring large-scale labeled datasets in challenging underwater environments characterized by high turbidity and limited accessibility. To overcome this data shortage, we constructed a synthetic dataset generation pipeline utilizing the Isaac Sim virtual environment. We implemented four distinct data augmentation stages tailored to mimic complex underwater variations: (1) a baseline simulating diverse natural habitats by placing 3D urchin models on complex rocky terrains; (2) an angle model adjusting camera pitch and distance to account for AUV instability; (3) a dense model mimicking clustered distribution patterns; and (4) a turbid model utilizing physics-based parameters to model light attenuation and murky visibility conditions. These augmented synthetic images were subsequently transformed into photorealistic underwater images using trained CycleGAN generators. The transformed images and their corresponding, automatically generated labels were paired to construct a highly robust training dataset for a YOLOv8-nano object detection model. Rigorous evaluations on challenging real-world test images demonstrated that the proposed multi-condition Sim-to-Real pipeline synergistically bridges the domain gap. The fully augmented model achieved a mean Average Precision (mAP50) of 71.3%, representing an exceptional 9-fold increase compared to the baseline's 7.8%. Crucially, the model successfully suppressed false positives and detected urchins in highly turbid environments, laying the essential groundwork for fully automated AUV-based sea urchin removal operations.

Keywords: AI/ML for oceanographic and environmental data analysis, Real-time decision systems for ocean robotics and autonomous operations, Foundation models and large-scale training for marine applications
Abstract: Underwater image segmentation remains challenging because changes in water conditions introduce color attenuation, scattering, contrast loss, and illumination degradation that adversely affect downstream visual perception. This paper presents a domain-adaptive Multi-GAN framework for seabed vegetation instance segmentation under heterogeneous underwater conditions. The framework integrates MobileNetV3-Small for visual-domain classification, domain-specific CycleGAN models for image enhancement, and YOLOv8n-seg for downstream instance segmentation. Three representative domains are considered: green coastal water, blue water, and low-light conditions. The experimental dataset is based on 303 real green-water seabed images collected in the Iwakuni Sea, Japan, with corresponding blue-water and low-light variants generated through controlled transformations. On a 70-image mixed-domain evaluation set, the classifier achieved 100% accuracy with an average standalone inference time of 4.69 ms per image. Domain-specific enhancement increased mean UCIQE and UIQM across all three evaluated conditions. Ablation experiments showed that mask mAP50 improved from 0.0036 for raw inputs to 0.1830 using a single CycleGAN and to 0.6841 using the proposed Multi-GAN framework, with a corresponding mAP50:0.95 of 0.3924. The complete pipeline processed 70 images in 36.58 s, corresponding to approximately 1.91 images/s. These results demonstrate the effectiveness of domain-aware enhancement for improving underwater instance segmentation and provide a modular framework for seabed monitoring and marine robotic perception.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Foundation models and large-scale training for marine applications, Data fusion, uncertainty quantification, and predictive modeling
Abstract: Autonomous Surface Vehicles (ASVs) operating in dynamic ocean environments face challenges from wave-induced motion that degrades sensor performance and navigation reliability. Accurate sea state estimation is critical for maintaining operational effectiveness and ensuring mission safety. This paper presents a proof-of-concept multi-modal framework for sea state estimation that combines vision-based semantic segmentation with inertial measurements. We train an IMU-based classifier that achieves 58.4% accuracy under leave-one-recording-out cross-validation across 19 field recordings, with 97.4% of predictions falling within one sea state of ground truth. We fine-tune a SegFormer model for maritime water segmentation, achieving 96.22% mean IoU on the LaRS validation set (91.25% at the early checkpoint used for deployment), and propose a mean maximum water extent (MMWE) metric for translating segmentation output into sea state estimates. We present a late-fusion architecture that integrates these modalities and demonstrate the complete system in real-time sea trials aboard a Havoc Rampage ASV. The results establish that both IMU motion signatures and segmentation-derived water extent carry sea state information, motivating their combination in a multimodal system as additional training data becomes available.

Keywords: AI/ML for oceanographic and environmental data analysis, Technology for precision fish farming, Real-time decision systems for ocean robotics and autonomous operations
Abstract: Understanding properties of fish populations has grown increasingly important over time. Analyzing fish length, width, and mass is integral to maintaining a healthy stock in the aquaculture industry, while counting wild fish populations fuels marine conservation and coastal reef restoration efforts. Current deep learning-based segmentation models offer a promising means to obtaining such information en masse. Most of these models, however, require costly and time-consuming per-pixel segmentation labels for training, which can take up to 2 minutes per fish to annotate manually. To address this inefficiency, we propose a weakly supervised segmentation pipeline that only requires point labeled data (1 manual click per fish) to produce accurate segmentation masks. Our pipeline consists of three stages. First, a point-supervised tracker predicts point annotations per frame. Then, points are converted into motion-informed language expressions. These language expressions are used to guide a language conditioned segmentation model (LMPM++) to capture temporal consistency and produce candidate segmentation masks. Finally, optical flow and background subtraction are applied as post-processing filters to suppress noise from incorrect segmentations. Preliminary results tested on the DeepFish dataset show significant improvement in foreground mean intersection-over-union (mIOU) compared to baseline models. Ultimately, we demonstrate the potential for a low-annotation fish analysis tool for aquaculture, population monitoring, and marine conservation.

Keywords: AI/ML for oceanographic and environmental data analysis, Benchmark datasets, reproducibility, and open-source frameworks, Citizen science and community involvement
Abstract: Marine biodiversity monitoring is constrained by the cost and logistical difficulty of repeated field surveys, while divers, tourists, guides, and researchers increasingly generate large collections of underwater imagery that could provide complementary ecological observations. This paper presents OASIS (Ocean Analytics for Species Identification and Surveillance), a computer-vision framework for screening opportunistic underwater images for conservation-relevant marine species. OASIS uses a hierarchical multilabel classifier that combines a ResNet-152 image representation with auxiliary taxonomic and visually observable attributes, distribution-balanced loss, and class-specific decision thresholds. The system was developed using 9,815 images spanning a 41-species output space and evaluated on a held-out test set of 1,465 images. The final model achieved a micro-averaged precision of 0.824, recall of 0.781, micro-F1 of 0.802, macro-F1 of 0.809, and exact-match accuracy of 77.3%. Species-level analysis revealed substantial variation in precision-recall behavior, including both recall-limited and false-positive-prone classes. Controlled ablations showed that the visual-cascade architecture produced modest improvements over a flat classifier, while distribution-balanced loss shifted the operating point toward higher recall. These results support the use of hierarchical multilabel classification as a screening layer for converting distributed underwater imagery into candidate biodiversity observations while retaining expert validation before ecological use.

Keywords: Benchmark datasets, reproducibility, and open-source frameworks, Data fusion, uncertainty quantification, and predictive modeling, Sensor synergy
Abstract: This paper presents CRAD, a manually annotated camera-radar alignment dataset for maritime sensor fusion, designed to support frame-level evaluation of camera-radar alignment, with manually verified roll and pitch references and an auxiliary effective sensor-height parameter. Camera-radar fusion on autonomous surface vessels requires frame-level alignment under vessel motion. However, obtaining reliable reference labels for frame-level alignment parameters in maritime scenes remains challenging. To address this challenge, CRAD provides manually verified roll and pitch reference labels, together with an effective sensor-height parameter used during projection annotation. Labels are generated with a custom projection-based annotation interface that jointly inspects image-domain and radar-domain cues. CRAD contains 599 labeled frames selected from 987 candidate samples across daytime RGB and nighttime IR sequences, enabling direct evaluation of frame-level alignment-parameter estimation and camera-radar projection.

Keywords: Optical sensors and adaptive optics, Marine optics technology and instrumentation, Oceanographic instrumentation and sensors
Abstract: Recent advances have expanded the use of event cameras (Event-based Vision: EVS, DVS, Event camera) in biological observation. EVS is a sensor that asynchronously outputs only brightness changes at each pixel. Compared with frame-based sensors (FBS), which acquire all pixels at fixed intervals, EVS offers high temporal resolution, low data rates, and reduced sensitivity to motion blur and exposure saturation. These properties make EVS suitable for capturing fast, non-stationary biological motion and the behavior of small particles. Applications have been reported in wildlife, insect, fish, and bat monitoring, as well as high-speed motion analysis. In marine science, EVS-based plankton observation has also shown promise. Microscopic organisms such as plankton and larvae are known to exhibit rapid movements of their swimming organs. With conventional frame cameras, particle tracking may be possible, but detailed measurement of the motion characteristics of individual particles is difficult. In contrast, EVS can not only count particles accurately but also capture periodic shape changes associated with swimming motion at high speed and extract motional features through frequency analysis. This enables classification of particles into active particles, which exhibit swimming motion, and passive particles, which do not. Such classification has often relied on post-sampling observation, but EVS enables direct in situ classification without water sampling. Because organisms show characteristic motion patterns depending on size and species, analysis of multiple motional features may also enable coarse taxonomic classification. This further suggests the potential of biomass measurement without water sampling. In this study, we conducted both laboratory observations using EVS and underwater in situ observations using an EVS installed in a pressure-resistant housing, demonstrating its effectiveness as a foundational technology for advancing EVS-based observation of marine particles and organisms.

Keywords: Imaging and vision, Holography and 3D imaging, Integration with autonomous platforms and observatories
Abstract: Coral reef health is increasingly threatened by climate-driven changes in ocean temperature and chemistry. Ocean acidification, the warming of the ocean, causes thermal stress and promotes coral bleaching and infectious diseases. By modeling the nature of coral reefs over time, scientists can better analyze and characterize ecosystem dynamics and predict long-term reef degradation. Satellite imagery is commonly used to study coral reef health, however, remains limited in temporal fidelity and spatial coverage in extreme environments. Direct observation and census-based monitoring are often used to ground-truth satellite data, but these approaches are largely qualitative, relying on human interpretation of 2D imagery and in situ surveys underwater. Because of this, efforts to consistently quantify coral reef characteristics over time are more difficult, particularly given the high dimensionality of the data and the multitude of interacting biological, environmental, and geometric factors involved. More advanced technology has been increasingly adapted, however robotic sensing remains challenging due to environmental unpredictability, poor visibility, and complex lighting conditions. Moreover, the non-rigid, continuously deforming nature of soft corals, makes the application of existing general-purpose reconstruction techniques unreliable. In this work, we address these challenges related to underwater dynamic reconstruction applied to coral reefs, and present some preliminary work to begin closing these gaps. We present a single-sensor light-field approach for modeling and reconstructing corals, inclusive of soft coral species, towards a more robust and deformation-aware framework to allow for quantitatively tracking of coral reef dynamics.

Keywords: Imaging and vision, Oceanography: physical, geological, chemical, biological, Integration with autonomous platforms and observatories
Abstract: The export of carbon to the deep sea occurs at the scale of individual sinking marine snow particles, and is controlled by biological processes that vary on timescales from seconds to years. Collectively, these particles export ~10 Pg of carbon into the deep sea each year, but monitoring this process is made difficult by the need to sustain high-resolution observations over long time periods. We developed a particle flux imaging instrument, the SINKing Ecology Robot (SINKER), to continuously monitor the flux of particles to the seafloor, quantify their ecological sources and sizes, and measure their sinking speeds. SINKER is moored to the seafloor just offshore of Monterey Bay, California, USA, and collects these measurements with 5 different camera systems, powered by an underwater cable. The data stream is continuously delivered onshore where images are processed through various automated pipelines. For sinking speed measurements, horizontal cameras inside the sample-collection tube collect images of particles at 1 Hz, and these data are used to detect and track particles as they sink. Upward-facing cameras image these same particles after they settle onto a collection plate, every 10 seconds. A wiping mechanism clears off the collection plate every 2 hours to prevent image occlusion. We will present the data from about ~10 months of data collection over the summers of 2025 and 2026, and demonstrate how sustained, high-resolution measurements of flux reveal the driving mechanisms of carbon export to the deep sea, and specifically in a dynamic coastal upwelling system. Specifically, we will present changes in particle size vs. sinking speed over time, and how this impacts total carbon export. We will also quantify the changing contribution of different particle classes to carbon export (e.g. fecal pellets from various animals and detrital aggregates). Details such as these (size, sinking speed, ecological source) are needed in next-generation models of the ocean’s biological carbon pump. Observations made by in situ particle imagers such as SINKER are providing the data needed to accurately monitor and model these processes across appropriate time scales and with realistic driving mechanisms.

Keywords: Systems and observatories, Oceanographic instrumentation and sensors, Oceanography: physical, geological, chemical, biological
Abstract: The EMSO LO (Ligurian Observatory), located in the Western Ligurian Sea, is a central component of the Laboratoire Sous-marin Provence Méditerranée (LSPM), a deep-sea research facility enabling continuous, real-time observation of the marine environment. As part of EMSO-France and European initiatives, it provides long-term monitoring of the ocean from the surface to the deep sea using a network of sensors connected by electro-optical cables. Integrated within LSPM, EMSO LO combines oceanography, biology, geosciences, and particle astrophysics. A key feature is its co-location with the KM3NeT-ORCA neutrino telescope, creating strong synergies between environmental monitoring and fundamental physics. The observatory measures a wide range of parameters, including temperature, salinity, dissolved oxygen, currents, and biogeochemical variables, enabling the study of ocean circulation, ecosystem functioning, and deep-sea processes. It provides unique insights into the coupling between surface and deep waters, including deep water formation, particle fluxes, and the biological carbon pump. EMSO LO also supports biodiversity research through imaging systems, acoustic monitoring of marine mammals, and bioluminescence measurements. In addition, it contributes to geophysical studies such as seismology and sediment dynamics. By delivering high-frequency, long-term datasets, EMSO LO plays a key role in assessing the impacts of climate change on ocean properties and ecosystems. As an open-access infrastructure within the LSPM framework, it fosters international collaboration and multidisciplinary research. EMSO LO represents a new generation of ocean observatories, providing a powerful platform to better understand the deep ocean and its role in the Earth system.

Keywords: Oceanographic instrumentation and sensors, Deepwater development technology, Oceanography: physical, geological, chemical, biological
Abstract: Underwater vehicles are widely used today to perform various tasks and operations in numerous fields. The continuous research and development of such robots enlarged the range of tasks they can perform in underwater exploration, manipulation, offshore structures and other military and industrial applications. We are particularly interested in rock coring for scientific purposes with the use of a Remotely Operated Vehicle (ROV). Today, underwater drilling requires the use of a rig either deployed from a vessel, or mounted on a ROV. The rock coring drill is fixed inside a structure and always performed manually and perpendicularly to the seabed. The aim of this publication is to present a novel and innovative drilling method based on the automation of a robotic arm mounted on a ROV that will allow scientists to drill rocks and outcrops in various geologic environments. Another important advantage of being able to carry out drilling using this type of technique is that it also gives the possibility to target a specific desired sampling area, which cannot be done from the surface.

Keywords: Oceanographic instrumentation and sensors, Deepwater development technology, Oceanography: physical, geological, chemical, biological
Abstract: Sediment core samples are essential for investigating deep-sea ecosystems and geological environments. They provide valuable records that enable the assessment of present-day disturbances— such as those driven by climate warming and anthropogenic activities—within the context of past sedimentary archives. They also preserve evidence of geohazards, including seismic events and volcanic activity, thereby contributing to the reconstruction of their frequency and impacts. Ship-based coring systems allow the collection of long samples but they do not benefit from decimeter-level positioning or visual feedback required for complex seafloor environments. Traditional push coring carried out by ROVs is limited to only 30 cm from the seafloor. As part of the DeepSea’Nnovation project (Equipex+ / France 2030), this article presents the development of a new vibrating coring system designed for deep-sea research. Drawing on 20 years of pioneering experience by MBARI, this system is optimized to allow the Victor 6000 ROV to collect sediment cores from 1 to 2 meters below the seafloor.

Keywords: Energy-harvesting and long-duration monitoring solutions, Oceanographic instrumentation and sensors
Abstract: Here we present the development of the Ocean Supplied Instrument Power or Ocean-SIP device which attracts biomass to generate power while deployed below the ocean’s surface. The Ocean-SIP takes advantage of lipid-rich biomass present in the ocean to generate acetate, a preferred fuel for microbial fuel cells, using a bioreactor. The biomass is attracted to a light, captured, converted to a slurry, then fed to the bioreactor. We envision this system connecting to a microbial fuel cell in the future, providing a self-sustaining approach to power for ocean sensors and vehicles.

Keywords: Oceanographic instrumentation and sensors, New vehicles, methods, and technologies for midwater exploration, Environmental DNA/eRNA for biodiversity assessment
Abstract: Understanding the diversity and function of microbial communities in low-oxygen water columns is essential because microorganisms serve as the primary biological engines driving global climate and environmental models. However, standard water column sampling methods frequently introduce thermal shock and depressurization biases when returning samples to the sea surface prior to processing. To overcome these limitations, the Microbial Sampler Submersible Incubation Device (MS-SID) was developed, enabling autonomous in situ sample preservation for both filtration and incubation studies. Technological advancements to the MS-SID include an upgraded 100-port Helical Rotary Valve (HRV) capable of managing fifty filtration ports with paired tracer preservative ports, an integrated multi-sensor feedback loop to correct mechanical drift and ensure precise valve alignment, and an autonomous scripting architecture that eliminates the need for a constant communication tether. Four sea deployments near the Northeast Atlantic shelf break to depths of 2000m validated the mechanical features and autonomous software. Field results demonstrated successful autonomous execution of "Sample-Incubate-Preserve" scripts, verified in situ tracer mixing, and achieved sample volume recoveries within 10 milliliters of programmed values.

Keywords: Systems and observatories, Oceanographic instrumentation and sensors, Access, custody, and retrieval of data
Abstract: The EMSO Ligurian Sea - LO site, a key node of France’s EMSO-FR infrastructure and the European EMSO-ERICnetwork, has been instrumental in long-term oceanographic monitoring since 2007. Its data legacy began with the IL07 mooring (2007–2013) at the ANTARES site (DOI: 10.17882/47130), marking the first systematic observations in the region.

In 2014, the mooring was relocated and redesigned as ALBATROSS, which remains operational today. This system has produced multiple data sets (DOIs: 10.17882/95265, 10.17882/95264, 10.17882/95263, 10.17882/95262, 10.17882/99982), expanding the range of physical and biogeochemical parameters measured. Between 2014 and 2016, the Secondary Junction Box (BJS) node (DOI: 10.17882/42680) enabled seafloor measurements, while the Instrumented Interface Module (MII) (DOI: 10.17882/75839), added in 2019, further enhanced observational capabilities. In 2023, the Bathyfamily module—including the BathyBot and its docking station (DOI: 10.17882/97948)—introduced real-time, high-resolution benthic monitoring.

The site’s long-term data acquisition has involved numerous collaborators, necessitating robust data harmonization efforts. SEANOE (Ifremer) serves as the primary platform for data publication, ensuring each data set is assigned a DOI and made publicly accessible through user-friendly interfaces. This guarantees visibility, traceability, and compliance with FAIR principles (Findable, Accessible, Interoperable, Reusable). Meanwhile, the technical teams of OSU Pytheas are responsible for the curation of the data sets themselves, applying standardized formats and controlled vocabularies. Scientists then perform data qualification, validating measurements and ensuring the integrity of Essential Ocean Variables (EOVs).

Additionally, the Data Management Steering Group (DMSG) of EMSO ERIC coordinates data policies across the network, overseeing the implementation of a common data management plan to ensure consistency in metadata and interoperability. This collaboration between SEANOE, OSU Pytheas, and the DMSG ensures that data from EMSO Ligurian Sea West are scientifically rigorous, fully integrated into European and global observing systems, and accessible for future research.

Keywords: Autonomous underwater vehicles, Systems and observatories, Oceanography: physical, geological, chemical, biological
Abstract: Exploration of the deep sea is essential to better understand marine ecosystems, the carbon cycle, and the impact of human activities on the oceans. In this context, the semi- autonomous benthic robot BathyBot, deployed at a depth of 2,500 m in the Mediterranean Sea and connected to the EMSO cable-based observatory, is conducting real-time observations using its sensors, cameras, and analyzers. The LSCE and the CNRS Technical Division will integrate BathyProf into this system a micro-profiler designed to measure oxygen microprofiles at the sediment-water interface. These data will enable estimates of benthic metabolism, carbon fluxes, and early diagenetic processes in deep sediments. This article will first introduce BathyBot, followed by the characteristics and design of BathyProf. It will then describe the principle of oxygen measurements using micro-optodes, before presenting the initial tests and results obtained. Finally, it will highlight the scientific opportunities offered by this instrument, which is scheduled for installation in late 2026 and deployment in 2028. BathyProf will enable the study of the seasonal and spatial variability of sediment oxygen demand, while forming part of a multidisciplinary approach involving, in particular, the CNRS and IFREMER. It also serves as the precursor to a profiler capable of operating at depths of up to 6,000 m as part of the DeepSea’Nnovation project.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Buoy Technology, Applications in geohazards, offshore energy, carbon monitoring, and environmental observation
Abstract: High-resolution measurement of CO2 exchange at the air-sea interface remains a critical unmet need in both carbon-cycle science and marine carbon dioxide removal (mCDR) verification. Existing measurement networks are insufficient to capture the spatial and temporal variability of ocean carbon flux, limiting the fidelity of natural carbon-cycle models and the development of robust monitoring, reporting, and verification (MRV) frameworks for mCDR methods. DeepGreen Solutions (https://deepgreensolutions.org) has developed COBALT(SM) (Carbon Ocean-Based Atmospheric Linked Telemetry), a buoy-based system designed to measure atmospheric CO2 exchange with the ocean surface at hourly resolution. Each COBALT(SM) unit combines a Sofar Spotter buoy platform — selected for its established satellite telemetry infrastructure — with a Vaisala GMP252 CO2 sensor, chosen through comparative testing as the optimal balance of measurement precision and deployment cost. Buoys may be deployed in tethered, free-floating, or self-positioning configurations. For free-floating deployments, three drogue designs have been evaluated to enable buoys to track subsurface currents and circulate within ocean eddy systems. Carbon flux at the air-sea interface is derived from the product of wind velocity and the atmospheric CO2 spatial gradient parallel to the wind, expressed as ppm/h. A fleet of COBALT(SM) buoys can thereby map capture and release rates across extended ocean areas in near real time. Methodology validation is currently underway via land-based comparison with an eddy-covariance tower, the accepted gold standard for surface-atmosphere carbon flux measurement. Parallel analysis of NASA OCO-2 and OCO-3 satellite data using the same flux derivation approach has yielded results consistent with buoy-based observations, and suggests that ocean regions with active mesoscale eddy systems are disproportionately significant sites of air-sea CO2 exchange. Scaled across a networked buoy fleet, COBALT(SM) has the potential to substantially increase the spatial density of carbon-flux measurements at the air-sea interface — advancing both fundamental carbon-cycle science and the MRV infrastructure essential to credible mCDR deployment.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Citizen-science networks and participatory ocean observations, Open-source hardware, software, and community innovation
Abstract: Abstract— GoSense, developed by RoboNation, enables real-time environmental data collection and analysis through low-cost sensing systems. The GoSense program includes an air-based sensor, GoSense Air+, and a water sensor, GoSense Marine, as well as educational resources tied to national learning standards and a data sharing community. GoSense provides accessible, modular sensing solutions to students, educators, and citizen scientists to explore their environment.

Keywords: Marine life and ecosystems, Oceanographic instrumentation and sensors, Real-time decision systems for ocean robotics and autonomous operations
Abstract: The TECTUGA project adopts a novel approach to marine monitoring that combines animal‑borne cameras on sea turtles with autonomous surface vehicles (ASVs) coordinated by artificial intelligence. By integrating biologging, cooperative robotics, and machine‑learning–based data analysis, the project enables rapid, low‑impact ecological assessments of marine habitats, biodiversity, and anthropogenic threats.

Keywords: Microbial community structure and function, Marine life and ecosystems
Abstract: Biofouling is a costly problem for the marine industry, with untreated growth causing functional and financial problems to marine vessels, submerged structures, instruments, as well as equipment. A substantial amount of research exists surrounding the effects of biofouling occurring on surfaces immersed near the surface, as well as efforts to create effective management and control techniques. However, there is less data on fouling rates at depths below the photic zone (200 meters/ 656 feet) and how intense hydrostatic pressure can influence the community composition on established surface fouling organisms that are subjected to different depths/pressures. To investigate fouling at depth, data was collected during two field campaigns (September 2024 and May 2025) at a deep-sea sub-tropical location to determine the effects of deep sea pressure on biofilm and incipient (newly settled) fouling. During each effort, a total of 16 PVC panels (15 x 27.5 inches.; 38.1 x 67.31 centimeters) were immersed off a pier in the harbor for approximately three and six months to accumulate biofouling growth. Panels (n=4) were later submerged at different depths known as shallow (~300 meters/984 feet) and deep (~1,000 meters/3,280 feet) at an off-shore location for different periods of time (1 hour, 4 hours, and 12 hours). Visual assessments, live/dead stain, chlorophyll a, biofilm adhesion and microscope analysis were all completed to measure changes in microbial health before and after exposure to the changing hydrostatic pressure. Small changes were noted in biofilm and macrofouling organism health for the shallow depths and short immersion trials. However, significant declines in overall viability of biofouling were detected when panels and organisms were subjected to deeper depths (1,000 meters) for the 12-hour immersion. Effects to biofouling from the deeper, longer immersion events included responses such as tubeworm tube evacuation, bleaching of macroalgae, and metabolic activity in microbial cells. This data begins to give insight into the effects of prolonged exposure to deep water conditions on the behavior and responses of biofilm and early settlement fouling.

Keywords: Marine life and ecosystems, Offshore structures, Sustainable offshore industries and the evolving blue economy
Abstract: As marine artificial structures (MAS) proliferate, their interactions with cetaceans increase. While localized studies suggest MAS aggregate prey, whether these effects translate to broader distributional shifts over extended periods remains unclear. I synthesized a multi-century cetacean sighting database (1789–2026) in the Gulf of Mexico, integrating survey and opportunistic records with dynamic, temporally validated MAS infrastructure data. Using a hybrid modeling framework: Restricted Maximum Likelihood GAMs (N ≥ 30) and parametric GLMs (N < 30), I evaluated habitat suitability across dynamic oceanographic covariates and localized MAS exposure metrics. Temporal filtering revealed that for 60%–90% of raw sightings, the nearest structure was unbuilt or decommissioned at observation time, underscoring the need to account for historical infrastructure dynamics to prevent "ghost platform" artifacts. Models demonstrated that offshore pelagic deep-divers (Physeter macrocephalus, Kogia spp., Ziphius cavirostris) maintained strict separation from dense platform corridors, whereas inshore species (Tursiops truncatus) exhibited strong positive associations with structure-dense shelf waters (>100–400 structures within 15–50 km). Significant negative relationships with structure density emerged primarily in data-rich pelagic species supported by multi-covariate baselines (≥ 5 covariates). Structure density held notable predictor importance for Balaenoptera physalus (8.6%), Kogia sima (6.3%), and Steno bredanensis (5.5%). Partial response curves revealed acute behavioral avoidance: K. sima displayed the steepest negative effect (dropping >40 log-odds units at low densities), while B. physalus exhibited a steady decline in presence probability as structure density increased. Critically endangered Rice’s whales (Balaenoptera ricei) showed complete spatial absence within 5 km of active structures. While primary oceanographic drivers dictate broad cetacean distributions, anthropogenic structure density acts as a significant secondary driver of spatial avoidance for sensitive pelagic species. This framework directly informs marine spatial planning and offshore energy decommissioning strategies.

Keywords: Oceanography: physical, geological, chemical, biological, Oceanographic instrumentation and sensors, Systems and observatories
Abstract: This study presents the In situ Oxygen Dynamics Auto-sampler (IODA), an innovative instrument designed to measure microbial respiration and organic matter remineralisation directly within the ocean. Conventional methods based on water sampling and laboratory incubations are limited by artefacts related to pressure changes, temperature variations, and confinement, particularly in deep-sea environments. IODA enables autonomous, high-frequency measurements of dissolved oxygen dynamics in situ, preserving natural environmental conditions. A major improvement in the system design is the use of a borosilicate glass incubation chamber, replacing polycarbonate and silicone materials. This ensures chemical inertness, low gas permeability, and improved measurement accuracy, especially for low respiration rates. Deployed at approximately 2000 m depth in the Mediterranean Sea, the instrument provides continuous observations of oxygen consumption, allowing direct estimation of microbial respiration in the bathypelagic zone. The results demonstrate that deep-sea respiration is highly variable over time, challenging the traditional view of a stable and weakly active deep ocean. Calculated carbon demand suggests that microbial activity at depth cannot be sustained by particulate organic carbon flux alone, indicating additional contributions from dissolved organic matter and lateral transport processes. These findings reveal a strong connection between surface productivity and deep ocean metabolism. The IODA system represents a major advancement in ocean observation technologies, providing new insights into the functioning of the biological carbon pump and improving our understanding of carbon cycling in the ocean.

Keywords: New vehicles, methods, and technologies for midwater exploration, Linking omics data to physical, chemical, and ecological processes
Abstract: Mesopelagic biodiversity is a vital planetary resource that is a critical component of the biological carbon pump and plays an important role in ocean food webs. Many mesopelagic animals expedite the flow of carbon from surface waters to the deep sea through their feeding and daily migrating activities. In diel vertical migration, considered the largest migration on our planet, animals generally ascend to surface waters to feed at night and return to depth during the day. This typical migration pattern is observed through the vertical movements of deep sound scattering layers. Complementary targeted environmental DNA (eDNA) sampling has the potential to provide the species-specific composition and depth distribution patterns of migrators comprising these layers. Here, we present a novel multi-vehicle approach termed “verified directed sampling” to sample eDNA precisely in and around deep sound scattering layers. The approach entails communication between the uncrewed surface vehicle DriX and the autonomous underwater vehicle Mesobot to direct sampling with the Oceanic-WHOI eDNA sampler based on acoustic backscatter detected by DriX and interpreted by shipboard and shore-based scientists. We successfully demonstrated this approach in an expedition on the E/V Nautilus in the Geologist Seamounts region near Hawaii in 2023. Verified directed sampling provides advantages over vertically integrated net sampling and standardized eDNA vertical transects with Niskin bottles that do not account for patchy and dynamic biomass, and is a valuable tool for understanding dynamic midwater ecological phenomena such as diel vertical migration.

Keywords: Biological and ecological processes in midwater environments, Diel vertical migration and trophic interactions, Midwater biodiversity and ecosystem structure
Abstract: Bruce H. Robison, Kim R. Reisenbichler, Rob E. Sherlock Monterey Bay Aquarium Research Institute robr@mbari.org reki@mbari.org robs@mbari.org The WEDGE: Long-term effects of expanding oxygen minimum zones on midwater communities Oxygen levels are decreasing ocean-wide, particularly in regions like the California Current, where oxygen minimum zones (OMZs) are expanding. One way to predict the ecological consequences of such changes is to conduct comparative studies, and with regard to oxygen profiles in the water column, the Gulf of California today looks a lot like what the central California coast may look like in 50 years. A long-term time series of quantitative, midwater ROV and AUV video transects, coupled with measurements of oxygen concentration, shows a distinct upward expansion of hypoxic water over the Monterey Submarine Canyon, which is linked to changes in the daytime depth centers of several mesopelagic species. From in situ measurements of oxygen consumption rates we can determine the critical partial pressure of O2 below which a species cannot regulate, and must compensate by moving, either closer to the surface, where the water holds more oxygen, or downward below the OMZ. Profiles of changing O2 concentrations and the depth ranges of target species in Monterey Bay show that as the OMZ expands, the vertical ranges of many species shift upward. Likewise, some species who are not diel vertical migrators, appear to shift downward into higher concentrations of oxygen. Investigations of counterpart species in the Gulf of California provide evidence to forecast what the future holds for Monterey Bay. By comparing the midwater communities of Monterey and the Gulf, we can see that an expanding OMZ acts as a wedge, shifting some species (but not all) from their historically preferred depths. Projected large-scale effects include: 1) fragmentation and re-combination of the mesopelagic community, 2) decreases in the abundance of established forage species, and 3) a shift to include more hypoxia-tolerant top predators.

Keywords: Midwater biodiversity and ecosystem structure, Autonomous underwater vehicles, AI/ML for oceanographic and environmental data analysis
Abstract: Observing mobile pelagic predators in situ remains challenging due to the scale of their habitats, behavioural responses to disturbance, and limitations of conventional sampling. Ship-based and remotely operated approaches provide limited coverage and can bias observations, constraining efforts to link predator distributions and behaviour with oceanographic structure. We present Piscivore, an autonomous visual observation system developed at the Monterey Bay Aquarium Research Institute (MBARI) to attract and observe pelagic predators and their prey with environmental context. Integrated with long-range autonomous underwater vehicles (LRAUVs), Piscivore enables adaptive, multi-week missions across dynamic features such as fronts and oxygen minimum zones. The system combines forward- and rear-facing high-definition cameras with a trailing passive attractor that promotes investigative behaviour without bait or active acoustics. Cameras are optimized for low-light conditions and synchronized with environmental sensors measuring temperature, salinity, dissolved oxygen, chlorophyll fluorescence, and optical backscatter. Machine-learning pipelines process large video datasets, using convolutional neural networks to detect and classify organisms, often to species level, and to quantify encounter rates and behavioural responses. Initial deployments document diverse pelagic assemblages, including fishes, sharks, gelatinous taxa, marine mammals, and seabirds, as well as multi-species encounters and predator-prey interactions rarely resolved with traditional methods. Coupled biological and environmental data reveal associations with productivity gradients and water-mass structure. Piscivore supports persistent, minimally intrusive sampling and contributes to emerging autonomous observing systems that integrate biological and oceanographic measurements, advancing the study of pelagic ecosystems.

Keywords: Aquaculture engineering, Industry-academia partnerships and translational technologies, Instrumentation, sensors and autonomous systems for sea-based aquaculture
Abstract: As the global demand for seafood increases, aquaculture operations need technology to improve efficiency. Bottle Upweller Systems (BUPSYs) have gained popularity for their more uniform growth rates and operational advantages compared to traditional systems cite{ver_design_1995}. In a BUPSY, oyster seed is placed in a bottle, known as a silo, with openings at the bottom and top. Nutrient-rich water is pumped in from below, causing seed to fluidize and suspend in the water column. As flow directly determines nutrient delivery, proper control is vital. As nurseries scale up, the burden of manually adjusting flow across large arrays of silos becomes increasingly time-consuming, motivating the need for automation. This paper presents the design and testing of a vision-based system that controls BUPSYs flow. A camera livestreams bed behavior to a computer, which quantifies spat motion and bed expansion using a dense optical flow algorithm and color thresholding. This data commands position changes for a valve actuator. Testing used two silo geometries to evaluate generalization across shape: a straight cylinder, which showed the most promise in early testing, and a hipped cylinder, the only commercially available geometry From a packed-bed state, both silo geometries reached fluidization in under 35 seconds and maintained fluidization over 90% of the time thereafter. Both systems recovered proper fluidization under pump disturbance and detected abnormal bed behavior. These results demonstrate the feasibility of a vision-based automated BUPSY, offering a framework for more consistent operation, reduced manual labor, and a platform for further experimentation previously infeasible in commercial settings.

Keywords: Hydrography / Seafloor mapping / geodesy, Marine geology and geophysics, Marine optics technology and instrumentation
Abstract: The Monterey Bay Aquarium Research Institute (MBARI) has partnered with 3D at Depth, now part of Kraken Robotics, to develop the second generation of Wide Swath Subsea Lidar (WiSSL2) for cm-scale mapping on the deep seafloor. The WiSSL2 has a single head that scans in a circular disk pattern, rather than 2 scanning heads, and together with a smaller electronics can, the WiSSL2 has an 80% reduction in size, weight and power than the first generation WiSSL. It still is a pulsed time-of-flight lidar with a 520 nm (green) laser, and the data are co-registered with 400 kHz multibeam sonar, stereo photographic cameras and strobe lights, DVL-aided INS, and a CTD for sound velocity. All components are rated to 6000 meters depth.

The sensors are mounted on an articulating frame to keep them normal to the imaged plane, from level seafloor to vertical canyon walls and hydrothermal chimneys. This package is currently operated as a toolsled on a ROV, and we are designing a hover-capable, 6 degree of freedom AUV to carry the payload. The vehicle is navigated autonomously, and flown at 0.2 m/s with a 3 meter standoff from the imaged surface to generate bathymetry at 5 cm lateral resolution from the multibeam sonar, 1 cm from the lidar, and 3 mm/pixel in the imagery. Photomosaics are quantitatively analyzed for observations of animals using a machine-learning algorithm trained on MBARI’s many decades-long dataset of ROV video. Areas of 120 m x 120 m can be covered in about 8 hours.

As mounted, the WiSSL2 directs about 120° of its scan at the seafloor, achieving a 10 m wide swath at a 3m standoff, which is a 50% increase in swath width compared to the original WiSSL. The range performance remains the same: the new lidar achieves a maximum range of about 10 m in moderately clear water, and tracks the seafloor well at a 3 m standoff when the water is just turbid enough to obscure the seafloor in the ROV video and the strobe-illuminated LASS photography. Results of mapping with the new WiSSL2 and modified LASS will be presented.

Keywords: Protection of offshore energy and critical maritime infrastructure, Security and resilience of AUV/USV/ROV platforms, Marine safety and security
Abstract: As the use of autonomous underwater vehicles (AUVs) grows due to the expansion of marine industries, the likelihood of incidents that result in uncontrolled or lost vehicles will increase. This creates the need for new technologies capable of non-destructively disabling and recovering underwater vehicles. A survey of patents and published solutions identified 20 non-destructive mitigation technologies that were classified into three categories: barriers, nets, and entanglement systems. These categories are evaluated using six qualitative criteria to identify deficiencies in existing mitigation technologies. The survey reveals that existing solutions are predominately passive systems with significant limitations in deep-water deployment and retrieval capabilities. Finally, expected directions for future research and novel solutions are discussed.

Keywords: Security and resilience of AUV/USV/ROV platforms, Maritime domain awareness and autonomous monitoring, Rapid-deployment systems and modular instrumentation
Abstract: Remotely Operated Vehicles (ROVs) are versatile assets for underwater exploration, research and security operations. When configured with specialized sensors and manipulation tools, operators on the surface can inspect and interact with objects from a safe standoff distance, making ROVs vital for detecting and neutralizing explosive threats. However, this standoff range is limited by the maximum length of the cable, which supplies power to the ROV and transmits data between the vehicle and operator’s workstation on the surface. The goal of Tethered Relay Buoy (TRB) is to increase the ROV’s operational standoff range by implementing a wireless radio link between the buoy and the surface command workstation. While this wireless link can increase the operational range in addition to the physical cable between relay buoy and ROV, it also increases the complexity in communication link. The TRB prototype was successfully tested in the bay and at sea to characterized the system’s baseline metrics and demonstrated it’s potential, though operational range performance was highly dependent on sea state conditions.

Keywords: Surveillance, detection, and threat-assessment techniques, Pollution monitoring, Search-and-rescue technologies and emergency response
Abstract: The maritime industry is progressively adopting low-carbon propulsion technologies to comply with increasingly stringent greenhouse gas reduction targets. Liquefied natural gas (LNG) has emerged as the most widely deployed transitional marine fuel. As LNG-fueled ships are increasingly operated in industrial ports, evaluating the thermal radiation hazards associated with accidental LNG pool fires is essential for safety management. This study proposes a configuration-based consequence assessment framework for open-deck LNG release scenarios. Three representative accident configurations are considered: (i) an unconstrained marine spill, (ii) a deck-confined spill determined by deck layout, and (iii) a coupled deck–marine spill caused by overflow from the deck. Radiation distances and exposed areas are estimated using a solid-flame model to support safety distance determination and risk-informed safety assessment.

Keywords: Ocean economic potential, Ocean energy, Ocean policy, governance, and regulatory frameworks
Abstract: An evolving and rich portfolio of novel ocean-based technologies offers opportunities for advancing coastal resilience and resource security. However, questions around scalability and commercialization remain unanswered for most of these innovations. Common barriers to upscaling include high capital costs, limited access to grid infrastructure, regulatory complexity, and investor uncertainty around these early-stage technologies.

One strategy for overcoming these obstacles is the development of marine “innovation centers” where technologies can be co-deployed and de-risked using shared resources and infrastructure. The marine innovation center model opens opportunities for assessing synergies between technologies, complementary energy and material flows, waste-to-value exchanges, and other avenues towards co-development and commercialization. Although several ocean innovation centers have been proposed or are in early development, there is no standardized, data-driven approach for evaluating locations and selecting technology mixes that best leverage existing infrastructure and are feasible at scale.

To address this gap, we present a multi-criteria decision analysis (MCDA) framework that can be used to assess site suitability, optimal technology combinations, and techno-economic implications at existing and proposed innovation hubs. By scoring a consistent set of criteria, the MCDA highlights where existing facilities and targeted infrastructure investments can be leveraged for the development of marine innovation centers.

Results from initial case studies will be presented including established marine innovation centers like the Hawaii Ocean Science and Technology Park, the largest desalination plant in the Western Hemisphere (Carlsbad, CA), a grid-connected wave energy testbed in Oregon, and other sites of interest. These results are relevant to the OCEANS audience because they provide a roadmap for upscaling nascent marine technologies, prioritizing sites where marine innovation ecosystems already exist. Our framework offers a data-driven approach that can be used to guide decision-making and commercialization strategies for researchers, innovators, and policymakers in the ocean technology space.

Keywords: Citizen science and community involvement, Oceanographic instrumentation and sensors
Abstract: Mediterranean Spy is a citizen science project, within the Ocean Spy application, to study deep-sea ecosystems in the Mediterranean Sea. It allows citizens to actively contribute to scientific research by analyzing images collected from the deep-sea benthic ecosystem. These images were taken at a depth of 2420 meters at the EMSO-LO deep-sea observatory in the Mediterranean Sea, near the French coastline. At EMSO-LO, a deep-sea robot known as BathyBot functions as a multi-instrument, Internet-operated vehicle that is remotely controlled in real time from shore. This benthic crawler collected images and videos on the seabed, recording the presence of marine organisms and environmental conditions during its 9-month deployment in 2023. However, the large volume of images and videos substantially exceeds the capacity of human experts to analyze and interpret the data manually. Through the Mediterranean Spy project, volunteers observe the images and identify visible organisms or object within 13 different categories. By classifying the most frequently observed groups of organisms or objects, participants contribute to advancing scientific understanding of biodiversity in deep-sea environments. In future studies, the resulting dataset will also support the development of artificial intelligence systems capable of automatically recognizing marine taxa on the basis of this training set. However, these opportunities are accompanied by the requirement to rigorously validate the annotations and to systematically analyse them in order to derive robust and meaningful scientific information. Moreover, the project enables researchers to monitor temporal changes in marine ecosystems.

The Mediterranean Spy project was launched in April 2024, and by August 2026 it had reached more than 23263 annotations involving 482 participants from 26 countries. From this dataset, we analyzed some pattern from the involvement of citizen. In an ecological point of view, the annotated dataset allowed to represent the dynamic over time of the listed organisms and objects. We are convinced that sharing such experience could help raising awareness about the importance of observing the deep-sea.

Keywords: Protection of offshore energy and critical maritime infrastructure, Microbial community structure and function, Marine materials science
Abstract: Biofouling are a major problem for the maritime industry. These unwanted marine growth cause operational and structural issues on ships, scientific instruments, and other coastal and offshore structures. In this study, we combine two environmentally friendly antifouling technologies for a more robust fouling control approach. The synergy of foul release coatings and ultraviolet-c (UVC) radiation have proven successful in the marine environment. However, an optimized approach of UVC application needs to be investigated. Our results indicate that a routine, proactive UVC duty cycle is better at mitigating biofouling than a continuous and reactive UVC dose.

Keywords: Ocean energy, Offshore structures, Hydrodynamics
Abstract: Offset-column mass is one of the largest single cost drivers in semisubmersible floating-wind support structures, but conventional lightweighting strategies forfeit waterplane stiffness and degrade motion response. This paper presents a feasibility analysis for an alternative: hollowing the offset columns of an OC4 semisubmersible to host oscillating water column (OWC) chambers, so that the geometric step that removes mass from the offset columns simultaneously creates an air-capped internal water column that supplies pneumatic restoring and damping. To select a viable hollow geometry, a two-fidelity approach is used, followed by a RANS-VOF verification under a moderate sea state to quantify the resulting motion behavior. The most consequential finding is a regular-to-irregular response reversal: under monochromatic excitation, the lightweighted platform exhibits a modest heave penalty relative to the solid baseline, but under broadband irregular excitation at the same nominal sea state, it achieves a roughly 20% heave reduction. The viscous damping recovery in the CFD model relative to linear theory further amplifies this trend. A passive-orifice study at the same sea state maps a clear hydro-pneumatic operating envelope: a wide restrictive band sustains hydro-pneumatic coupling and useful pneumatic power, while an over-open limit collapses chamber pressure and breaks the compensation mechanism. The combined evidence supports OWC-integrated lightweighting as a credible direction for cost-effective stabilized FOWT platforms and motivates future chamber-specific active airflow control within the identified envelope.

Keywords: Buoy Technology, Hydrodynamics, Ocean energy
Abstract: Access to safe drinking water remains a critical global challenge, particularly in coastal regions affected by natural disasters, where infrastructure and electricity are often disrupted. Although seawater is abundant in these environments, converting it into potable water typically requires energy-intensive desalination systems that depend on stable power sources. This project develops and evaluates a floating wave-powered system that uses a two-body wave energy converter (WEC) with a mechanical power take-off (PTO) to amplify low-speed wave motion into mechanical power and drive an axial piston pump for pressurized, unidirectional flow. A Berkeley wedge is incorporated to reduce vortex formation and improve hydrodynamic performance. Time-domain simulations using WEC-Sim were conducted to study how wave period, mooring stiffness, and PTO damping affect power output, guiding design choices. A 1:8 prototype was designed, fabricated via 3D printing, assembled, and tested in a controlled wave tank to quantify pumping behavior and identify failure modes that informed subsequent design iterations. Using modeled power and conservative reverse osmosis (RO) specific-energy assumptions, a full-scale unit (~2.0 m × 1.25 m) is projected to produce >400 L/day of freshwater, enough to supply ~200 people during the “golden 72 hours” after disasters.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Vehicle design, Rapid-deployment systems and modular instrumentation
Abstract: Abstract — Oceanographic research is frequently hindered by a market gap between cost-prohibitive enterprise surface vessels and payload-limited micro-USVs. To address this barrier, this work presents the design, methodology, and validation of the All Purpose Low-cost Uncrewed Surface Vessel (APLUSV). Engineered to maximize scientific versatility, the APLUSV is a highly modular mid-sized catamaran constructed with commercial off-the-shelf components modified with minimally specialized manufacturing. The vessel features a structural semi-monocoque deck and a frame constructed of aluminum tubing with universal mounting features to offer a total of 7.87 square meters of mounting space, that can support up to 82 kilograms of diverse payloads. The collapsible mast and thruster mounts enable convenient transport in a pickup truck. The vessel utilizes an open-source autonomy architecture equipped with differential thrust and redundant Beyond Visual Line of Sight (BVLoS) compatible telemetry. Field trials validated APLUSV’s high-precision autonomous navigation, reliable propulsion, and an extended mission duration. Ultimately, the APLUSV democratizes accessibility for marine data collection.

Keywords: Aquaculture engineering, Remotely operated vehicles, Vehicle design
Abstract: Aquaculture serves as a sustainable alternative to land-based agricultural systems. However, they struggle to match the output of their terrestrial counterparts due to several factors. In oyster aquaculture, one such factor is the heavy reliance on manual labor. Floating systems typically require the flipping of bags or cages to clear biofouling, which restricts water and nutrient flow into the containers. There can be thousands of these bags and each can weigh up to 30 kgs (60+ lbs). MIT Sea Grant has developed a prototype unmanned surface vessel (USV) for use in flip-bag oyster aquaculture systems to remove this reliance on manual labor. Several versions have been made, each building off of the results of the previous iterations' results. Using findings from the previous vessels, the third generation of the Oystermaran makes improvements on oyster field maneuverability and flipping ability. In this paper, we describe the overall design and fabrication of the system and the results of field testing. The system has been tested in MIT Sea Grant and the Charles River, demonstrating considerable progress over previous generations in terms of hydrodynamic performance and structural integrity while flipping baskets.

Keywords: Vehicle navigation, Vehicle performance, Autonomous underwater vehicles
Abstract: This paper proposes a disturbance observer-based model predictive control (DOB-MPC) framework for precise trajectory tracking of unmanned surface vehicles (USVs) under environmental disturbances and model uncertainties. While MPC can explicitly handle system constraints and underactuation, its performance degrades in the presence of disturbances and modeling errors. To address this limitation, the proposed approach incorporates disturbance estimates into the prediction model, ensuring disturbance-aware and feasible control actions while avoiding instability caused by actuator dynamics. The proposed method is validated through open-sea experiments using an 8 m class USV, demonstrating significant improvements in trajectory tracking performance compared to nominal MPC.

Keywords: Real-time decision systems for ocean robotics and autonomous operations, Vehicle performance, Automatic control
Abstract: Hydrofoiling sailing vessels present a strongly nonlinear control problem in which aerodynamic, hydrodynamic, and inertial forces are tightly coupled and environmental conditions vary continuously. This work evaluates classical and learning-based control strategies for a hydrofoiling catamaran under variable model and environmental conditions, using a validated six-degree-of-freedom nonlinear simulation. Two controllers are compared: a Linear Quadratic Regulator (LQR) and a Proximal Policy Optimisation (PPO) reinforcement learning agent. Under nominal conditions, LQR demonstrates superior tracking performance with more stable actuation effort. However, Monte Carlo analysis across 5000 trials shows that PPO maintains stable foiling flight in 67% of cases, compared to 12.6% for LQR. A delay-aware PPO variant trained under 250 ms actuation delay retains stability in 63% of trials in the delay environment, compared to 3% for the standard PPO, with no architectural modification required. These results demonstrate that while LQR provides efficient control near its design point, the substantially greater robustness of PPO to environmental variability and hardware constraints makes it the more practically deployable controller for vessels operating across a broad and uncertain envelope, where robustness is a more appropriate design objective than local optimality.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Open-source hardware, software, and community innovation, Dynamic positioning
Abstract: This work presents the development of an active heave compensation (AHC) system for an existing fiber optic tether system and targets the sea conditions faced during the tether breakage of a past ROV deployment. The system adapts the active heave compensation approach intended for offshore cranes and implements it onto an affordable off-the-shelf microcontroller with estimation, prediction, and control methods. The estimation system produces attitude estimation from sensor fusion of measured IMU data and uses adaptive filtering to estimate the heave. The prediction algorithm using auto-regressive modeling and a Levinson recursive least squares algorithm is used to produce a forecasted heave and velocity. This is then successfully used as a pure feedforward command to drive compensator action from the motor to correct for heave. Simulation and bench tests are used to demonstrate the performance of the system.

Keywords: Foundation models and large-scale training for marine applications, Benchmark datasets, reproducibility, and open-source frameworks, Sonar imaging
Abstract: Acquiring large-scale annotated sonar datasets for underwater perception is prohibitively expensive due to the high cost of marine operations and the limited availability of controlled test environments. We developed FURo-Sim, a GPU-accelerated, physics-based sonar simulation framework built on Unreal Engine 5 (UE5) that generates physically realistic synthetic sonar imagery with pixel-accurate ground-truth annotations at negligible marginal cost.

FURo-Sim couples UE5's GPU rasterization backend with an acoustic model based on the sonar equation, supporting forward-looking sonar (FLS) and side-scan sonar (SSS) modalities. Target strength is computed via a roughness-dependent BRDF combining diffuse Lambertian and specular Phong lobes; transmission loss models spherical spreading and frequency-dependent absorption; multipath propagation is handled through recursive GPU ray tracing. Three independent noise sources (speckle, crosstalk, ambient) are composited to reproduce the stochastic appearance of real sonar returns.

A key contribution is automatic multi-modal ground-truth generation: per-pixel 2D segmentation masks, 2D bounding boxes (highlight-only and highlight-with-shadow variants), semantically labeled 3D point clouds, and 3D oriented bounding boxes are produced alongside every rendered frame without manual labeling. The simulator also includes 6-DoF AUV and ASV vehicle models, an auxiliary sensor suite (RGB camera, IMU, GPS, pressure sensor, magnetometer, LiDAR, DVL), and ROS 2 / Python interfaces.

Quantitative validation against real DIDSON 300m and Oculus M750d imagery — using MSE, PSNR, MS-SSIM, and LPIPS metrics — confirms that each acoustic component contributes meaningfully to realism and that higher-order reflections improve structural fidelity in complex scenes. FURo-Sim provides a complete, end-to-end simulation and dataset generation platform for field and underwater robotics research where real annotated data remains scarce.

Keywords: AI/ML for oceanographic and environmental data analysis, Sonar imaging, Ocean exploration
Abstract: Bathymetry measurement and estimation is critical for understanding the structure of underwater environments. While multibeam echosounders (MBES) provide accurate depth measurements, their deployment on surface vessels limits spatial resolution in deep water. Sidescan sonar (SSS), in contrast, can be deployed on autonomous underwater vehicles (AUVs) to operate closer to the seafloor, enabling high-resolution acoustic imaging. However, the inherent ambiguity in measurements and the limited observation coverage of sonar datasets degrade the performance of state-of-the-art methods for underwater bathymetry estimation from SSS. In this paper, we propose a differentiable neural rendering framework for SSS that enables effective estimation of bathymetry from SSS measurements. Our approach incorporates a measurement model based on SSS geometry, explicitly accounting for its wide elevation aperture and associated ambiguities. We evaluate our method on both synthetic and real-world datasets, demonstrating improved robustness and accuracy compared to alternative approaches.

Keywords: Passive and active acoustic monitoring, Anthropogenic noise characterization and mitigation, Sonar signal processing
Abstract: Single-hydrophone Doppler analysis of ship tonals often supports several speed modes. In 18 waterjet craft pass-bys, a permissive Doppler screen nominated the low-speed region for 10 passes, although only 6 were commanded low. We test tonal–pedestal contrast (TPC) as a gate criterion for applying a low-speed restriction to the Doppler set, rather than as a speed estimator. TPC compares a tracked ridge with its local broadband pedestal and is aggregated as a ratio of temporal medians. The six commanded 10-kn/2000-rpm passes had higher mean TPC than the rest (Welch p = .0014). Among the other 12 passes, Doppler nominated low for four and the TPC criterion was met for two different passes. With each period held out, a midpoint AND gate allowed contraction for all 6 targets and none of the 12 others. For the six contracted passes, the median candidate-set measure changed from 46.0 to 22.5 km/h. Across all 18 passes, reference inclusion remained 16 of 18. The smallest TPC decision margin was +0.13 dB.

Keywords: Acoustical oceanography, Signal coherence and fluctuation, Sound propagation and scattering
Abstract: This paper examines tide-influenced semi-diurnal travel-time perturbations along the 166 kilometer path from the Kauai Beacon (KB) to the ALOHA Cabled Observatory (ACO). The path lies close to the Ka'ena Ridge, one of the world’s most energetic baroclinic tide-generation sites. For 72 hours, starting December 4th, 2025, at 00:00 UTC, the KB duty cycle was raised from 2% (normal operation) to 33% (20-min transmissions every hour). This time frame coincides with a full moon (December 4, 2025, 23:14 UTC) and brackets the maximum tidal forcing (spring tide). The travel time perturbations for each transmission were calculated using the integrated autocorrelation phase at one period lag (IAP1) method, which uses the phase of the full received field of each transmission (rather than individual peaks in the arrival pattern) to resolve perturbations. IAP1 was used because strong bottom interactions along the KB-ACO path (all ray paths are bottom-interacting) produce variable multipath structure that hinders tracking individual arrivals over time. IAP1 results yield a clear semidiurnal signal with an M2-fit amplitude of 7.9 ms with a phase of 55.66° (phase value relative to beginning of time series), confirmed by an independent envelope cross-correlation estimate (8.1 ms, 54.75°). A barotropic prediction derived from a Princeton Ocean Model simulation of the M2 tide accounts for 1.7 ms of this amplitude (phase 17.55°), and a BELLHOP ray-trace through a ROMS sound-speed field, which captures a portion of the sound speed variability (including baroclinic and mesoscale temperature and salinity variability), contributes 1.4 ms (phase 69.46°). The combined model contributions have an amplitude of 2.8 ms and a phase of 46.43°. The difference between the combined model and the observed data is roughly 5.1 ms in the semidiurnal band. The unresolved signal leaves room for further investigation of tidal energy at the location. The results show that the KB–ACO path, combined with IAP1 processing, is a practical tool for near-field observation of the Hawaiian Ridge tide, complementing basin-scale tide studies and results from the Hawaii Ocean Mixing Experiment (HOME).[ONR Ocean Acoustics program, grant no. N00014-24-1-2546.]

Keywords: AI/ML for oceanographic and environmental data analysis, Foundation models and large-scale training for marine applications, Benchmark datasets, reproducibility, and open-source frameworks
Abstract: Accurate benthic characterization is essential for coral reef monitoring, yet existing image-based approaches have largely focused on coral cover, which provides only limited insight into reef condition, ecological resilience, and functional composition. In particular, coral growth-form composition constitutes an important ecological dimension because different morphologies contribute differently to recovery dynamics, structural persistence, and habitat provision. To address this gap, we developed a morphology-aware semantic segmentation framework for pixel-level classification of seven canonical coral growth forms in underwater benthic imagery, including branching, corymbose, digitate, massive, encrusting, foliose, and submassive corals, together with background. The model was trained and evaluated using 3,024 manually annotated orthomosaic patches generated from high-resolution underwater surveys conducted in Kenting National Park, Taiwan. UNet++ was adopted as the backbone network because its dense skip connections and multi-scale feature fusion are well suited to the delineation of fine colony boundaries and spatially adjacent coral morphologies. The proposed framework achieved an overall mean Intersection over Union of 0.833, an overall accuracy of 0.988, and an F1-score of 0.879 on the test set. Performance was highest for branching and massive corals, whereas lower accuracy for digitate, encrusting, and submassive classes reflected greater morphological ambiguity and stronger visual similarity to surrounding substrate. These results demonstrate that semantic segmentation of coral growth forms can move reef-image analysis beyond coral-cover estimation toward ecologically richer benthic characterization. The proposed framework provides a practical basis for scalable, morphology-resolved reef monitoring and offers a foundation for future integration of structural complexity metrics into unified ecological assessment workflows.

Keywords: AI/ML for oceanographic and environmental data analysis, Model-based signal processing techniques, Hydrography / Seafloor mapping / geodesy
Abstract: Forward-looking sonar (FLS) is widely used on autonomous underwater vehicles (AUVs) for seafloor mapping and collision avoidance. However, it usually has a wide vertical aperture. As a result, elevation information is unavailable because echoes returned from different elevation angles at the same range and azimuth will merge into a single image pixel. This will make terrain reconstruction challenging. To this end, this paper presents a three-stage convolutional neural network to predict per-pixel elevation angles from a FLS image for 3D cloud generation. The network is trained using co-registered multibeam echosounder (MBES) measurements. The staged decomposition is designed to perform no-return rejection, partial-return clas- sification and elevation regression. It is motivated by the fact that 86.5 % of pixels in an FLS image have no valid returns. The elevation regression stage is trained with a locally weighted Huber loss function normalized by the total-variation and Laplacian smoothness terms along each beam. The method is validated using the simulated data generated in the Stonefish simulator. The pipeline attains an F1 score of 86.3 % on the classification (stage 1&2), an RMSE of 1.03 deg (0.018 rad) in simulation and RMSE of 6.51 deg (0.11 rad) deg on the field dataset. To make the trained model applicable to field data, speckle noise was applied on the simulated sonar images and bilateral filtering was applied with intensity thresholding to enhance the simulated images that are used in model training. Then the trained model is applied to the field data to create a point cloud of the terrain and compared with the results collected using a Norbit MBES mounted on an ASV.

Keywords: AI/ML for oceanographic and environmental data analysis, Oceanographic instrumentation and sensors, Pollution monitoring
Abstract: The rise in plastic pollution in our oceans and rivers necessitates automated systems for microplastic detection. Scout is a recently developed impedance-based microplastic sensor that can detect and distinguish microplastics from other particles commonly found in waterways. This work evaluates signal processing and machine learning techniques for analyzing Scout’s data output to detect particles passing through the sensor and classify those particle detections as plastic or non-plastic. Peak detection, Short-Time Fourier Transform (STFT), and wavelet transforms were tested for particle detection. Six machine learning classification models were evaluated to differentiate plastics from biological materials and bubbles. Detection results showed wavelet-based detection as the best performing (93.7% recall) with peak-based detection having acceptable performance (87.9% recall) with a much smaller compute time. Results indicated that convolutional neural networks (CNN) achieved the highest classification accuracy (98%), but random forest models offer a highly efficient alternative (93% accuracy) with significantly lower inference time, making real-time microcontroller processing for an impedance-based microplastic sensor theoretically feasible.

Keywords: Autonomous underwater vehicles, Open-source hardware, software, and community innovation, Oceanographic instrumentation and sensors
Abstract: Cylindrical watertight enclosures are widely used across various underwater systems, from unmanned underwater vehicles (UUVs), to remotely operated vehicles (ROVs), to various sensor platforms. However, electronics are typically built on rectangular PCBs arranged in horizontal stacks, which inefficiently occupy the circular cross-section volume that is critical for both payload capacity and buoyancy management. This paper presents CUB (Central Utility Board) and CougarTail, a general-purpose system designed to address this gap. CUB is a circular PCB sized for 4-inch-diameter enclosures that consolidates a Raspberry Pi Compute Module 5 (CM5) and an STM32 microcontroller, while also providing power management features and auxiliary connections. Mounted coaxially, CUB reduces the electronics stack of our CougUV from 200 mm of tube length and 709 g to 25 mm and 156 g, returning that length and mass budget to payload and buoyancy trim. CougarTail is an open-source companion sensor mast that houses a GPS antenna and two dual-band (2.4 and 5 GHz) omnidirectional PCB antennas. Both components are validated through bench testing and integration on a CougUV platform, our small open-sourced torpedo UUVs.

Keywords: Low-cost, scalable oceanographic sensors and platforms, Energy-harvesting and long-duration monitoring solutions, Oceanographic instrumentation and sensors
Abstract: Maritime sensing faces challenges from lack of energy and communications in remote regions. Especially when constrained by cost, sensors struggle to provide effective coverage of the maritime environment in ports and waterways. This work introduces a standardized, small, low-cost Shore Pod sensor platform for deployment in large numbers to autonomously monitor the maritime domain. Using this platform, we evaluate solar energy harvesting for moored Shore Pods. Given historical irradiance data of a target deployment environment and onboard sensor characteristics, we propose and evaluate an energy buffer sizing process to maximize uptime while minimizing volume, mass, and cost. We identify three energy-scheduling pitfalls and evaluate operating strategies to avoid these pitfalls and maximize uptime across operating environments. Top-performing Shore Pod operating strategies achieve 98.9% of maximum uptime.

Keywords: Machine-learning classification of acoustic signals, Passive and active acoustic monitoring, AI/ML for oceanographic and environmental data analysis
Abstract: Underwater acoustical signal classification became a matter of significance over a century ago when submarines became threats to national security amidst World War I. Since then, our understanding of the depths of the ocean has expanded, yet we are barely scratching the surface. If classification could be done without the sending of signals, but rather just with a passive listening of the underwater soundscape, the rate at which our knowledge of the ocean expands would accelerate dramatically.

The Monterey Accelerated Research System (MARS) lies 52 km off the coast of California and uses a hydrophone and three accelerometers to record raw acoustic data year round. For the purpose of this research, the data was divided into minute-long files. This data was initially processed into time-frequency representations of power spectral density levels and bearing. This research focuses on how these time-frequency representations can best be used for signal classification.

Recent research by Artorix de la Crus and Mae L. Seto indicated that Rényi entropy may be key in appropriate signal classification. Rényi entropy is a specialized form of Shannon Entropy which describes a measure of the unpredictability of an event.

In addition to entropy, other acoustic features were evaluated for their potential in signal classification: kurtosis, spectral flux, and spectral centroid. For each combination of acoustic features, unsupervised learning methods were applied and the classification results were compared. Initial efforts were made with KMeans Clustering and future efforts will implement additional methods to improve preliminary labeling. After removing features one by one, it was evident that when feature vectors included a measure of entropy, the clustering performed considerably better.

These results highlight the potential for Renyi entropy to advance the validity of unsupervised signal classification for providing preliminary labels to passively recorded data.

Keywords: Passive and active acoustic monitoring, Distributed hydrophone arrays and networked acoustic systems, Low-cost, scalable oceanographic sensors and platforms
Abstract: Accurate underwater localization and reliable ocean-to-land data transmission are essential for ocean monitoring and autonomous underwater vehicle operations. Existing solutions often suffer from either high complexity or limited localization robustness. This work presents a low-cost and low-complexity framework that integrates acoustic source localization with long-range wireless data transmission. For underwater localization, three passive hydrophones are employed with Generalized Cross-Correlation with Phase Transform (GCC-PHAT) to estimate the time difference of arrival (TDOA). To mitigate multipath-induced false peaks in reverberant underwater environments, a grid-based spatial evaluation method combines the correlation information from multiple hydrophone pairs to identify the most spatially consistent source location. Experimental measurements in an enclosed water tank demonstrate successful localization in six of seven tested positions. In parallel, an ocean-to-land wireless communication system integrating a buoy-mounted RF modem and an onshore software-defined radio (SDR) base station is developed. The proposed framework provides a practical and scalable solution for low-cost underwater sensing, localization, and long-range maritime data transmission.

Keywords: Integration with autonomous platforms and observatories, Sensor synergy, Autonomous underwater vehicles
Abstract: This work presents the development, implementation, and testing of a sensor-enabled towfish with the use of autonomous systems. The towfish is equipped with magnetic field sensors, an inertial measurement unit (IMU), and pressure sensors to allow for subsea magnetic field mapping. This research focuses on four primary goals: the characterization of the performance of the towfish and electromagnetic sensors, an assessment of the quality of the measured data, determination of the minimum length of the tether needed for sensor package isolation, and the assessment of the impact of towing on vehicle drag and its maneuverability. Tests are currently being conducted to assess the towing behavior of the towfish from an autonomous surface vehicle (ASV). Following successful tests on board the ASV, the towfish is being optimized for use on board the REMUS 100 autonomous underwater vehicle (AUV).

Keywords: Marine life and ecosystems, Oceanography: physical, geological, chemical, biological, Ocean stewardship, conservation strategies, and climate resilience
Abstract: Seasonal migrations and aggregations of baleen whales in coastal shelf waters of the U.S. Mid-Atlantic Bight (MAB) and Southern New England (SNE) represent regular occurrences in the regional marine ecosystem. However, the extent to which baleen whale migration and aggregation patterns vary seasonally and interannually, and how those distributions are driven by oceanographic conditions, remains largely understudied. This study examined seasonal and interannual variability in baleen whale occurrence and distribution during two oceanographically distinct years (2023-2024) using effort-standardized data collected from the National Oceanic and Atmospheric Administration (NOAA) Northeast Fisheries Science Center Northeast Region Right Whale Aerial Survey program. Additionally, a Hudson Canyon case study investigated environmental conditions associated with a documented multi-species whale aggregation observed during 2024. The baleen whale species observed during our study included fin (Balaenoptera physalus), humpback (Megaptera novaeangliae), North Atlantic right whale (NARW, Eubalaena glacialis), minke (Balaenoptera acutorostrata), sei (Balaenoptera borealis), and blue (Balaenoptera musculus). Environmental conditions within the Hudson Canyon were compared between 2023 and 2024 using satellite-derived oceanographic datasets, including sea surface temperature, sea surface temperature gradient magnitude, chlorophyll-a concentration, and bottom temperature. Whale distributions differed substantially between seasons and years characterized by contrasting oceanographic conditions. Seasonal dynamics were species-specific and varied between years. Winter of 2023 sightings were primarily concentrated around the midshelf 61% of the time, whereas spring, summer, and fall sightings occurred primarily in outer shelf waters, accounting for 50.7%, 77%, and 79.2% of the sightings respectively. In 2024, 67.2% and 96.3% of sightings in spring and summer, respectively occurred in outer shelf waters, while winter sightings were more evenly distributed across the depth strata, but showed an increase in nearshore occurrence compared to 2023. Compared to 2023, conditions during 2024 were associated with relatively colder shelf waters

Keywords: Seafloor engineering, Applications in geohazards, offshore energy, carbon monitoring, and environmental observation, Buoy Technology
Abstract: The Chumash Heritage National Marine Sanctuary contains diverse benthic habitats that remain undocumented due to the logistical barriers of deep-sea exploration. High-end autonomous underwater vehicles capable of reaching the sea floor require massive funding, and conventional benthic landers routinely leave metallic debris behind due to corrosion. To resurface, many submersibles use galvanic timed releases that corrode in seawater or heat-activated bolts to ensure weights are jettisoned even if the primary systems fail. These traditional fail-safes result in heavy-metal pollution and leaching.

The Deep Ocean Research-Explorer (DOR-E) addresses these challenges as an affordable, zero-trace autonomous benthic lander for 1000m biological surveys. To attract fish, a rigid vinyl-coated steel bait cage stays within the camera’s field of view and provides consistent scent diffusion. The camera is a Go Pro Hero 4 with a rechargeable battery able to record two hours of footage on one charge. Contained in an aluminum housing, it can be remotely activated with a small remote before deployment. This system is paired with a red-filtered LED light to illuminate the seafloor without disturbing light-sensitive deep-sea organisms.

A sand-based weight release system replaces corrosive fail-safes with bioinert beach sand. The system uses an electromagnet housed inside a pressure casing that allows an internal magnet to securely hold an external drop plate carrying the sand ballast. An autonomous timer de-energizes the circuit, dropping the sand to restore positive buoyancy without leaving metallic debris. 3D printed conical spike landing gear prevents the frame from becoming embedded in soft sediment.

Testing confirms the electromagnet safely holds 45 lbs of ballast through the casing, and FEA verifies the Aluminum 6061-T8 housing withstands 10 MPa of hydrostatic pressure. The sand design achieves recurring cost of only 3.05 per drop. By the conference, live deployments will provide videos of attracted carnivorous creatures to assess marine life population density. DOR-E aligns with the community’s focus on AUVs and remote sensing by providing a sustainable, highly scalable blueprint for deep- ocean observation.

Keywords: Ocean energy, Offshore structures, Data fusion, uncertainty quantification, and predictive modeling
Abstract: Design and assessment of ocean energy infrastructure requires robust knowledge of the site metocean conditions influencing power production and device survivability. Realistic subsurface flow velocities are of particular relevance to phenomena that are highly sensitive to directionality; yet, directionality changes in current with depth have historically been neglected due to a lack of accessible measurement data. To address this gap, we seek to leverage modern reanalysis data within a staged statistical learning workflow to provide representative joint-probability bins of metocean data, including standard wind and wave features with the addition of multi-directional current profiles, for the Humboldt Bay Wind Energy Area. Through the use of complex empirical orthogonal functions, 98.4% of the variance in current profile data is represented within a reduced-parameter space using three modes of spatial distributions. This reduced parameter representation of currents with depth allows us to deploy well-established clustering methods on a tractable eight-dimensional dataset of wind, wave, and current conditions. We use a modified maximum dissimilarity algorithm to generate multi-dimensional bins of joint metocean conditions. Results of this work can be directly used to study phenomena that are sensitive to directionality, making the workflow a valuable contribution to the ocean energy and ocean data fields.