United States

Maritime vessel classification is an important part of ensuring the security and safety of shipping, transport, and other maritime traffic, as well as facilitating the management and surveillance of various ocean regions. This importance has motivated many research studies which propose various techniques and models for vessel classification, mostly based on data collected through the Automatic Identification System (AIS). AIS datasets consist of collections of AIS pings which contain information about a specific vessel including its type and dimensions as well as position and heading, among other metadata. Studies using these datasets take many different approaches to structuring and processing AIS data and use various types of learned models for classification. Due to the high volume of available data, none of the existing works use a standardized dataset and often report classification results with different target classes and data representation, making them difficult to compare. In this work, we investigate how the structure and representation of AIS data affect a vessel classification task. We primarily divide existing approaches into two groups, (1) sequence-based methods and (2) feature based methods. Sequence-based methods keep AIS data in its sequential format, classifying vessels based on a window of continuous pings. Feature based methods rely on curated feature vectors containing statistical aggregations of data over a given time window. Within sequence-based methods, we also identify several encoding strategies for representing individual AIS pings. We implement versions of each of these approaches and evaluate them to answer the following research questions: 1. Are there notable performance differences between sequence-based and feature-based methods? 2. When using sequence-based methods, what is the impact of different encoding strategies when using the same model architecture? 3. Are static features alone (the ship dimensions reported in an AIS ping) enough to classify vessels without the need for trajectory information? By investigating these research questions, our research helps to unify and categorize many existing works, identifying how the representations of AIS influences classification.

NAML 2026

An Evaluation of AIS Data Representations for Maritime Vessel Classification

vessel classification

data representation

machine learning

benchmark

### Welcome to the 10th annual workshop on Naval Applications of Machine Learning!

NAML is the premier annual event showcasing current machine learning research relevant to Naval applications. We welcome participation from the defense community, industry, and academia. The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team. 
Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.
#### On-Demand Access

* This is an on-demand site and all the recorded content from the workshop is available here. 
* Access to the recordings is exclusively available to registered participants.
* Each registrant has received an email notification, with instructions on how to access the content.

[![](https://assets.underline.io/markdown_image/1/image/60cd12aa376c191d3467ea2ee60944ea.png)](https://naml2026.org/)

The goals of NAML are:

* To spread awareness of current machine learning research relevant to Naval applications
* To connect machine learning researchers with experts in military requirements
* To build and strengthen collaborations within the defense research community

#### *"What we do as a Navy is a human business. Don’t pass up an opportunity to come together as a team. Don’t pass up leadership opportunities, and don’t outsource them to a PowerPoint. If you have an opportunity to come together as a team, to increase belonging, do it."*
Admiral John M. Richardson, 31st Chief of Naval Operations
<br><br>

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Please log in using the email address you registered with.** If you have any issues logging in, please contact us at <naml2026@underline.io>.

Let’s celebrate the past decade by investing in the next. 🚀


Please log in to explore this event.

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Explore the latest advancements in machine learning tailored to Naval applications 🔎**

With on-demand access to NAML 2026, you can watch all lectures & browse poster presentations at your convenience, while also engaging directly with presenters via the Underline platform - ask questions, exchange ideas, and stay connected.

**If you still haven’t registered, you can do so now.**
👇
**Take advantage of a special one-time offer: full access for just $99, with complimentary registration for students, government, and military employees.**

**Get access here: https://naml2026-ondemand.paperform.co/**

Please check your spam folder, or contact our support team at **<naml2026@underline.io>** if you registered, but still haven't received an access email.

Let’s celebrate the past decade by investing in the next. 🚀

Please register!

The 10th Annual Workshop on Naval Applications of Machine Learning (NAML 2026) was held in San Diego, California, March 2-5, 2026. NAML is a forum for interdisciplinary experts from government, industry, and academia to discuss results of ongoing research on theory and applications of machine learning to address some of the most challenging problems faced by our naval forces. 



technical paper

Accurate underwater object recognition remains challenging due to strong backscatter from the target, environment, and water column, as well as the limited use of polarization in conventional imaging systems. Most existing RGB cameras and LiDARs employ single or dual linear polarizations, leaving the rich interaction between light polarization and submerged objects underexploited. In this work, we present an multi-polarization LiDAR imaging and recognition pipeline that combines novel hardware with turbidity-aware computation. On the hardware side, we develop a metasurface-based LiDAR capable of transmitting and receiving four polarization states spanning linear and circular bases, yielding 16 transmit–receive polarization pair images per acquisition. On the algorithmic side, we introduce a turbidity-aware subset selection (TASS) framework that first performs turbidity-aware normalization (TAN), estimating a scene turbidity index from Stokes statistics (e.g., degree-of-polarization decay and return variance) and reweighting each polarization pair to equalize effective SNR and stabilize polarization structure. Lightweight scores are then computed that jointly capture contrast-to-backscatter ratio, polarization preservation (DoLP/angle-of-polarization stability), and inter-pair diversity (low mutual information). A simple greedy policy selects a budgeted subset of k polarization pairs, which are fused by a compact CNN for recognition. We evaluate our system on a submerged dataset of 10 alphanumeric targets imaged under three discrete turbidity levels and varied geometric configurations. Experiments show that (i) multi-polarization fusion significantly outperforms single-polarization baselines, (ii) models trained on optimized polarization subsets can surpass using all 16 pairs, and (iii) TASS with TAN improves accuracy over naive selection and matches or approaches the performance of full-pair fusion, while being more compute- and acquisition-efficient. Overall, our results demonstrate a practical, budget-conscious path toward turbidity-robust underwater object recognition without modifying existing multi-polarization LiDAR hardware.

Metasurface-Based Polarimetric LiDAR and Turbidity-Aware Fusion for Underwater Imaging

The Naval Innovative Science & Engineering (NISE) project, Localization of Aircraft using Several Shipboard Optical Sensors (LASSOS) addresses the critical need for autonomy and perception in degraded environments to maintain our fleet’s competitive advantage. LASSOS aims to estimate the position of incoming aircraft relative to ships using only Pan-Tilt-Zoom (PTZ) cameras, supporting both flight deck tracking to reduce sailor cognitive load and enable Unmanned Aerial Vehicle (UAV) recovery operations. While current commercial off-the-shelf solutions rely on active sensors like Light Detection and Ranging (LiDAR), Navy operations require Emissions Control (EMCON)-compatible passive sensing solutions.
Our technical approach comprises three key tasks: (1) developing machine learning algorithms for PTZ camera calibration, (2) implementing aircraft detection and tracking using object detection models to maintain aircraft within camera field of view, and (3) combining this information to achieve accurate 3D aircraft localization relative to the ship.
The first task, PTZ camera calibration, has two components: an offline calibration algorithm, and an online calibration algorithm. In the offline algorithm, a dataset is captured of several images from each camera containing a set of calibration patterns at different pan, tilt, and zoom values. This is used to estimate and optimize the intrinsic and extrinsic calibration parameters for each camera. The online algorithm is then used to estimate the calibration parameters in real time based on each camera’s current pan, tilt, and zoom value. The second task tracks aircraft using a YOLO-family object detection model. Once an aircraft is identified, the cameras automatically center the aircraft within its view to maintain a high-quality track. The third task takes the estimated calibration parameters from the first task, and the bounding boxes from the second task to provide a 3D position estimate of the aircraft.
This work directly addresses a critical capability gap for NAVAIR, as several UAV recovery systems currently under investigation require precise positional data of incoming aircraft. By providing this data through passive means, LASSOS enables these recovery systems to function in EMCON environments. Beyond recovery operations, the passive sensing technologies developed through LASSOS can be applied to numerous naval perception challenges, advancing autonomous systems across the fleet.

Localization of Aircraft using Several Shipboard Optical Sensors (LASSOS

Optical Character Recognition (OCR) for complex scripts such as Chinese and Arabic remains a persistent challenge, particularly in handwritten contexts where character variation, stroke irregularity, and lack of standardized spacing significantly degrade accuracy. Tesseract is a widely used open-source OCR engine that supports over 100 languages and provides pretrained models, but its out-of-the-box performance struggles with these handwritten texts. Although Tesseract offers robust recognition for many printed scripts, baseline error rates for handwritten Chinese and Arabic often exceed 60-70%, limiting its applicability in operational environments. 

This work focuses on improving recognition accuracy for these scripts by fine-tuning Tesseract models using Tesstrain, an open-source toolkit that enables custom dataset integration, domain-specific adaptation, and retraining of Tesseract’s language models. This research was conducted as part of the development of POLAR (Portable Optical Language Analysis and Recognition), a mobile data-capture and translation application designed to support U.S. Marine Corps operations. 

To evaluate performance, we constructed pipelines that normalize prediction and ground-truth labels, then computed Character Error Rate (CER). For each sample, the image, prediction, CER, and metadata are written to structured CSVs, supporting analysis, threshold scoring, and model comparison. Using publicly available datasets, we retrained the base Tesseract Chinese and Arabic models to better capture the characteristics present in handwritten text. 

Testing shows substantial accuracy improvements after fine-tuning. The fine-tuned Chinese model demonstrates a clear reduction in handwritten CER compared to the baseline model, with a CER dropping from 79% to 28% against our test dataset. Similarly, the fine-tuned Arabic model CER drops to 22% from 61%. These results highlight Tesstrain’s effectiveness in adapting Tesseract to challenging scripts and writing styles that its default models do not handle well. 

Recognizing the tradeoff inherent in handwritten-optimized models, we also evaluated performance on printed text. The fine-tuned Chinese model experienced a notable increase in printed CER, while the fine-tuned Arabic model showed modest degradation. This highlights the need for task-specific model selection or dual-model strategies when deploying OCR across mixed handwritten and printed domains. 

Overall, this work demonstrates that with targeted training data and evaluation, open-source OCR models can be significantly improved for handwritten text in traditionally difficult scripts. These enhancements were implemented within the POLAR platform, enabling seamless OCR processing directly on Marine Corps mobile devices. The methods require modest computational resources, are fully reproducible, and can be extended to other languages, making Tesstrain-based fine-tuning an effective approach for enhancing OCR performance.

Reducing OCR Error Rates for Chinese and Arabic Handwritten Text Using Tesseract + Tesstrain

Visual anomaly detection identifies content that deviates from a known distribution. This remains a major challenge for modern vision models, such as the Vision Transformer (ViT). ViTs excel at classifying seen classes, but lack a principled mechanism to flag truly novel content. 

We propose using hyperdimensional computing (HDC) to solve the problem of visual anomaly detection. HDC captures information into a hypervector (HV), offering a mathematical representation of human memory. For example, we might have a set of hypervectors that represents the visual appearance of a hammer. Each of these HVs would have 10,000 or more dimensions. HVs can be compared to determine how related each of the concepts are to each other. In our hammer example, we can create a mean class to represent the average HV for a hammer. We can compare tools against this to determine their similarity. High similarity indicates the concepts are strongly related, low similarity might indicate that this is another class of tool (i.e., it is out-of-distribution).

Visual anomaly detection has a wide variety of Naval applications. In this work, we consider the application of visual anomaly detection to shipboard robotics. Visual anomaly detection can be used in a wide variety of scenarios including locating rust on bulkheads, leaking mechanical components, frayed wires, or other tasks such as finding debris or locating spills. 

We create HVs from pre-trained networks, such as a ViT. To create a hypervector, the image is processed by the network. We ignore the classification label and instead use the intermediate features to build the HV. This approach is known as Hyperdimensional Feature Fusion (HDFF) (Wilson et al., 2021) and has been validated on problems like finding out-of-class data from CIFAR-10 as well as problems such as transfer learning and gesture recognition (Tran et al., 2025). 

We use HVs to enable visual anomaly detection on our robotics platform. Here, the robot can build a set of HVs representing the normal appearance of an environment. The observed appearance will then be compared to determine if it is visually anomalous. We demonstrate this approach on the problem of cleaning up spills, in which visual anomaly detection is confined only to the appearance of the ground. Although our results are preliminary, we believe that the success in this domain will translate to other areas of shipboard maintenance.

Hyperdimensional Vision Anomaly Detection for Autonomous Shipboard Robotics

Real-time perception in aerial and maritime domains is difficult because detection-plus-tracking pipelines depend on running detectors at every frame, making them slow and computationally expensive. To address this, we develop and present a suite of data-propagation strategies that allow detections to be reused or predicted across intermediate frames, enabling stronger detectors to run less frequently while maintaining accuracy and recall. Data propagation works by using previously computed detections and either copying them forward or updating them through geometric or appearance-based cues, effectively reducing detector load and increasing FPS. We study three cases: (1) Copy Strategy, which directly propagates prior detections; (2) Motion Strategy, which updates detections using a Kalman filter to estimate future positions; and (3) Copy-or-Motion with Re-ID, which selects the better propagation path based on feature similarity from a re-identification model. We evaluate these techniques on challenging aerial and maritime videos featuring diverse object sizes, motion blur, and dynamic backgrounds.
Our results show that no single strategy dominates across all conditions. Copy propagation struggles with small or fast objects but performs well for medium to large targets in slow-moving defense scenarios such as ships or aircraft. Motion propagation is more resilient to rapid movement but can drift with irregular motion patterns. Surprisingly, the combined Re-ID–based approach does not consistently outperform the simpler methods; reliance on feature matching increases latency and reduces the geometric advantages of motion modeling. Through qualitative and quantitative comparisons, we demonstrate how data propagation significantly boosts FPS while preserving detector quality. We also provide practical insights on selecting the appropriate propagation strategy based on object scale, movement characteristics, and application requirements, offering a foundation for building efficient real-time perception pipelines.


Data Propagation Techniques for Real-Time Tracking in Aerial and Maritime Scenarios

Ever since the first introduction of computers and physics-based models, the US Navy has utilized model and simulation (M&S) techniques to improve designs, develop new technologies, and increase the capabilities of the fleet. As M&S techniques have evolved, so have the applications within the naval community. For example, the introduction of real-time simulations and hardware in the loop testing allowed rapid prototype development and the de-risking of shipboard power systems. Presently, innovations in artificial intelligence (AI) and machine learning (ML) algorithms offer another immense opportunity for the US Navy, as AI/ML could greatly impact how warfighters complete their missions. 
Just like any new technology however, AI/ML must be thoroughly validated before they can be fully accepted. In the case of hardware in the loop testing, decades of experiments showed that the testing of hardware in a virtual environment is just as rigorous as traditional testing methods. Unfortunately, AI/ML does not have a standard deployment method. This creates a challenge in comparing different AI/ML solutions to a problem or evaluating the performance of an agent/algorithm in different scenarios. In this work, the authors propose one potential method by demonstrating the evaluation of machine learning algorithms via software in the loop (SIL) testing, where the algorithm connects directly to a real-time simulation. In our model, a supervisory controller for the battery energy storage system (BESS) commands different battery systems, telling them when to charge, discharge, or remain idle. Furthermore, each of four batteries are assigned to individual aggregated load center (ALC), representative of various loads one might encounter on a ship (e.g. radars, lights, motors, etc). Finally, each battery and ALC connect to a common dc bus through a dc-to-dc converter, creating a notional four-zone medium voltage dc architecture. Once trained, the BESS algorithm is deployed into the simulation using specialized hardware (i.e. GTSOC) to evaluate its performance. If the performance is unsatisfactory, the algorithm learns from it and improves itself. If the results meet all expectations, however, the algorithms can now be safely tested and deployed in physical hardware with less risk. The ease of deploying algorithms through SIL allows for the direct comparison between any number of ML algorithms to be done in a fair, cheap, and effective method. 


Evaluation of Machine Learning Algorithms using Software in the Loop Testing

Our adversaries are using low-cost unmanned drones, leveraging a highly adaptable and capable platform for a variety of tasks from intelligence gathering to precision strikes. These drones are transforming the traditional battlefield, offering military advantage at the fraction of the cost of traditional systems. From a counter-drone perspective, they pose a significant challenge because they are a constant changing mix of commercial platforms with modified payloads. Statically built AI models for counter-drone cannot keep pace.

Future conflicts for counter-drone will require the US Navy to address the agility of its AI-system with continuous model adaptation at the tactical edge. The AI and MLOps solution must be operational where communication is denied or degraded and where static, pre-trained models quickly fall behind evolving threats. Recent Navy efforts in Project AMMO/Overmatch have demonstrated the reduction of model fielding time from months to days, with the adoption of a robust MLOps workflow. Within this project, we have demonstrated model compression and speedup while maintaining model accuracy and while staying within SWaP constraints. Furthermore, we demonstrated real-time field updatable models with watermarks and encryption to secure against adversary exploitations.

Beyond Project AMMO and under Army Project Linchpin, we have further reduced the model update time to minutes, and with MLOps software embedded in a field portable tactical kit. Our Linchpin work points to a concept of operations where threat libraries and models evolve in near real time, and operational in networked denied environments. New data on enemy targets can be captured and labeled with AI-assistance, and new targeting algorithms can be tuned quickly. We leverage embeddings from foundation models to reduce overall model training time. This workflow is built for operators, and can serve as an adaptive counter-drone pipeline. We plan to provide a demonstration of this technology at NAML, as an approach to refresh counter-UAS, swarm, ISR, and electronic warfare models as fast as the threat evolves. Such a capability can lock in a sustained battlefield advantage that matches the SecWar emphasis on speed, adaptability, and AI-enabled decision dominance.


Continuous model adaptation at the tactical edge

High-Performance Computing (HPC) systems generate massive numerical simulation data that domain experts struggle to analyze efficiently, while Large Language Models (LLMs) excel at natural language but cannot interpret raw numerical outputs. We present a novel, domain-agnostic multimodal pipeline that bridges this gap by automatically transforming numerical simulation data into natural language, without requiring manual annotation, visualization, or expert labeling.

Our proposed pipeline includes a multi-modal model, obtained by fusing a state-of-the-art LLM (such as one from the Mistral or Llama families) with an encoding pipeline, which projects a numerical sample onto a meaningful latent space. The two components are fused by means of a cross-attention layer, which connects the textual and numerical information.

An integral part of the research is the training methodology, which includes the generation of representative data samples by means of numerical analysis tools. The data samples aim at linking the latent representation of the numerical data to quantities of interest for scientists, and their natural language descriptions.

As a proof of concept, we present an ongoing project, which is part of Destination Earth, European Commission’s flagship initiative aimed at creating an accurate digital twin of the Earth system. The project is a weather specialization of our suggested approach, fusing an existing AI model (AIFS, developed by ECMWF) with an open source LLM (from the Mistral family), to enable real-time analysis of numerical weather prediction data.

As this project is still in progress, we will highlight its achievements and the challenges we are currently facing.


Designing a multimodal LLM to analyze numerical data

Many US Government agencies have complex system diagnostic workflows that need automation. However, their systems are typically unable to support Large Language Models (LLMs) due to a lack of GPU hardware infrastructure. Adding GPUs to such environments is difficult and cloud providers are not viable due to DDIL requirements. Instead, they rely on IT experts to interpret system logs and remediate issues. This places strain on IT experts in mission-critical environments and can delay error resolution time. Existing automation methods rely on simple machine learning and brittle parsing pipelines that lack extensibility and adaptability to new log patterns without extensive re-engineering. These are limitations that LLMs overcome through few-shot learning and flexible reasoning.
We propose using a modular Knowledge Distillation (KD) framework along with diagnostic workflow decomposition to balance each agency’s adaptable automation needs with their hardware constraints. KD is a process in which a larger LLM is used as a teacher to train smaller and cheaper student models that will replicate the teacher’s capabilities. By working with Subject Matter Experts (SMEs) to decompose troubleshooting workflows into specific subtasks, we allow the teacher to adapt to the subtasks through prompting and few-shot learning. Next, the teacher generates labeled synthetic data to fine-tune each student model for a specific subtask. For tougher subtasks, additional KD techniques (logit KD, counterfactual distillation, etc.) are leveraged depending on the unique complexities of the subtask. The resulting student models can then be deployed in environments lacking GPU support and may be reused modularly across other overlapping troubleshooting workflows.
To verify the success of our approach, we work with SMEs to develop sets of sample input log data with associated acceptable resolutions for each workflow. We use evaluation techniques such as semantic similarity, LLM-as-a-judge, and domain-specific heuristics to assign quantitative evaluation values to the results. We first show that the teacher model achieves acceptable evaluation scores across an SME-identified troubleshooting workflow as a baseline. We then show that the student models achieve comparable evaluation scores when remediating that same error. For each subtask, we also use the teacher model to develop gold standard datasets of labeled test data to assess accuracy and precision of each student model individually.
Our work provides a path toward modular diagnostic frameworks that can be deployed securely in US Government environments that lack resources. Our framework enables SMEs to be more efficient by reducing their cognitive load when solving these complex problems and enables domain workers who are not yet “experts” (e.g. IT3 right out of IT school) to be more effective when approaching these problems. In both scenarios, labor cost and system downtime for problem resolution are therefore minimized.

Modular Knowledge Distillation for System Diagnostics on Edge Devices

The U.S. Marine Corps (USMC) faces significant challenges in adapting its operational frameworks to embrace modern technologies, notably in the integration of advanced data sciences and AI/ML. For example, maintenance cycle challenges include: 1) excessive administration overhead, Marines spend more time on data entry and status updates then hands on maintenance; 2) fragmented data sources, critical information is often scattered across multiple systems; 3) tribal knowledge, critical know how often resides with a few experienced maintainers rather than being documented; 4) manual re-entry errors, information that is copied between systems without automation often gets mistyped.
To leverage advanced data sciences and AI/ML, we research and implement a prototype tool called AI Data integration & Automation for Maintenance (ADAM). The expectation of ADAM is to 1) act as the data set SME, providing clear insights and ensuring all conclusions are based on established standards; 2) develop, analyze, and summarize Excel reports to provide clear insights that inform leadership decisions, resource allocation, and mission readiness; 3) perform trend analysis to identify issues and patterns, using insights to improve preventive maintenance, anticipate risks, and enhance equipment reliability.
In this paper, we will review and analyze methods to address these challenges and expectations as follows:
• Apply and compare natural language analysis methods including large language models (LLMs, Llama3:8B, Qwen3:30B, GPT-5), retrieval augmented generation (RAG), and lexical link analysis (LLA), extract patterns from unstructured data to characterize the risks. 
• Apply collaborative learning agents (CLAs) to analyze distributed and dynamic data without moving raw data to a central location by using a network of autonomous learning agents that collaboratively discover patterns, anomalies, and emerging concepts. 
• Utilize RAG severs and explore context and machine learning for the Global Combat Support System Marine Corps (GCSS-MC) to provide context-aware decision support and predictive analytics for logistics and maintenance.
• Design testing procedures for varying complexity designed to assess each model’s consistency, accuracy, reasoning, and prompt-following, while evaluating how performance changes with prompt quality with and without context.

We will demonstrate the system using the use case from the Electronics Maintenance Platoon, 11th Marines Regiment HQ Battery.

Acknowledgement: Thanks to Quantum Intelligence, Inc. and the SBIR contract N6833525G0023 “Developing Collaborative Learning Agents (CLA) System for the USMC Warfighting Systems, & Human Factors Integration.”

Disclaimer: The views and conclusions contained in this document are those of the authors and should not be interpreted as others or representing the official policies, either expressed or implied of the U.S. Government.


Premium content

Downloads

Next from NAML 2026

Metasurface-Based Polarimetric LiDAR and Turbidity-Aware Fusion for Underwater Imaging

Similar lecture

Anomalies in Automatic Identification System (AIS) Maritime Traffic Data