China Releases Vast Ship Detection Dataset to Train Naval AI

More Than 2,000 Image Pairs for Finding Ships at Sea

A Chinese military research team has released what it describes as the world's largest publicly available dataset for ship detection, pairing more than 2,000 visible light and infrared images with nearly 20,000 annotated instances. The collection, called the dual modal ship detection dataset, or DMSD, gives researchers material for training artificial intelligence to recognise vessels in difficult maritime conditions.

The research appeared in the peer reviewed Journal of Radars in January 2026. Scientists from the Naval Aeronautical University in Yantai, Harbin Engineering University and the Chinese Academy of Sciences' Institute of Computing Technology developed the collection and made it freely available to the global research community.

DMSD brings together two views of the same maritime subjects: the appearance recorded by an ordinary camera and the heat patterns recorded by an infrared camera. Training algorithms on these paired views is intended to improve recognition when darkness, glare, haze or other interference makes a vessel difficult to identify.

The potential applications include surveillance systems, drones and missiles. For those systems, knowing that a ship is somewhere nearby is only part of the task. Identifying the vessel and keeping track of its position as it moves are harder problems, particularly when radar is being jammed, deceived or otherwise degraded.

The release represents an investment in AI for naval warfare, with a focus on the recognition failures that can interrupt surveillance or targeting. Its scale also gives researchers and defence developers a larger pool of examples from which to build and compare detection systems.

The central idea is straightforward: a machine that has seen many carefully labelled examples of ships under different conditions has more material from which to learn what to look for when a new image arrives.

Why the Ocean Makes Recognition Difficult

Sunlight reflecting off water can produce glare strong enough to overwhelm an optical sensor. Rapid changes in weather, atmospheric distortion over long distances and cluttered backgrounds can also make images less useful. The ocean often provides few stable visual reference points to help a computer separate a vessel from its surroundings.

Ships add another difficulty because they keep moving. Their changing positions and orientations alter what sensors see, while movement can blur an image. Sea spray, fog and haze further reduce clarity. An algorithm trained mainly on stable scenes on land may struggle with this combination of motion, shifting light and a largely uniform water surface.

The observing platform can move too. A drone, aircraft or satellite sees a ship from a changing angle. This creates perspective shifts and parallax, the apparent movement of an object caused by a change in the observer's position. Those effects complicate the work of following a target from one observation to the next.

DMSD addresses these problems with annotated examples captured under different sea states and lighting conditions. The aim is to help AI learn recognition patterns that remain useful as the appearance of the sea and the vessel changes.

When Radar Faces Interference

Radar remains a powerful way to find ships, but electronic warfare can interfere with its returns. Jamming can disrupt the signal, while spoofing can create misleading information. Either can leave uncertainty about a vessel's location or identity.

Military researchers have consequently pursued sensor fusion, which combines information from different kinds of sensors. The approach gives a system more than one way to observe the same scene. If one source becomes less useful, another may still provide information needed to continue detection or tracking.

For autonomous weapons, recognition errors can have serious consequences. A system may lose a manoeuvring target during the final stage of an engagement or struggle to distinguish a warship from a civilian container ship. Training data that reflects changing maritime conditions is intended to reduce those failures.

DMSD supplies paired visible and infrared material for that work. Researchers can use it to develop algorithms that connect the detail available in a conventional photograph with the thermal information available from an infrared sensor.

What the Dataset Contains

DMSD pairs visible light images with infrared signatures of the same vessels under identical conditions. This lets a learning system compare how a ship looks in ordinary imagery with how it appears through its thermal emissions.

Long Gao, the study's corresponding author and a researcher at the Chinese Academy of Sciences, described the team's data collection process:

Data were acquired using [synthetic aperture radar or SAR], radar, visible light cameras, infrared cameras and other sensors mounted on shore based and airborne platforms.

The collection process used several kinds of sensors, while the released dual modal dataset centres on paired visible and infrared images. That pairing is the basis for learning relationships between the two kinds of observation.

The nearly 20,000 annotated instances provide labelled examples for supervised learning. In this method, researchers show an AI system both an image and the expected answer, allowing it to learn how particular visual patterns relate to a labelled object.

Each annotation includes precise bounding boxes and classification labels. A bounding box marks the area occupied by an object in an image. The labels support the development of finer discrimination between vessel types, operating states and orientations.

The volume and variety of examples also address a problem known as overfitting. With a small training collection, an algorithm can become too closely adapted to the examples it has already seen and perform poorly on unfamiliar images. A larger collection gives developers more varied material for training recognition models.

The dataset's combination of scale, paired imagery and detailed annotation makes it a shared resource for researchers working on the same maritime detection problems.

What Infrared Adds After Dark

Infrared cameras detect thermal radiation rather than relying on the visible light that ordinary cameras record. At night, when conventional images lose useful detail, thermal sensors can still register heat patterns from a ship's engines, exhaust systems and hull against cooler water.

That capability gives infrared imagery a particular role in maritime surveillance and targeting. It can remain useful when visual camouflage or poor visibility makes conventional identification difficult. The dataset provides examples intended to help AI use those thermal patterns alongside ordinary images.

Infrared detection also operates passively. Radar sends out electromagnetic signals and measures what comes back, whereas an infrared sensor observes radiation already coming from objects. It can therefore collect information without broadcasting a radar signal that reveals its own activity.

This passive operation makes infrared sensing harder to detect and leaves it outside the traditional electronic jamming methods aimed at active radar. Its value grows when a military system needs another observation method during radar interference.

Visible light and infrared bring different strengths to the same task. Conventional imagery can provide fine detail for classifying a vessel when lighting is adequate. Infrared can support detection through darkness, haze and light fog when the visible image becomes less useful.

Training on both views is intended to help systems maintain a track as conditions change or countermeasures affect one sensor. The objective is continuous observation, with information from one type of sensor supporting the other through transitions that might otherwise cause contact to be lost.

For autonomous weapons operating under time pressure, this continuity matters because engagement decisions may be made without human intervention. The paired approach is intended to support the switch between sensor types while preserving the information needed for targeting.

A February Claim Near the Strait of Hormuz

Iran's claim in February that it had successfully struck the aircraft carrier USS Abraham Lincoln near the Strait of Hormuz illustrated the military stakes surrounding maritime targeting. Washington rejected the claim and said the missiles or drones involved had not come close to their intended target.

The episode brought attention to the distance between sending a weapon toward a naval force and guiding it accurately onto a moving vessel. Positive identification and terminal guidance, the final steering needed as a weapon approaches its target, remain demanding tasks in a crowded maritime setting.

The Strait of Hormuz carries nearly a third of global LNG and oil shipments through narrow waters. Commercial traffic, nearby shorelines and environmental interference create a complicated scene for any system attempting to select and follow one ship.

Accurate recognition has two connected purposes in such conditions. It helps a military system maintain contact with the intended target, and it helps separate military vessels from civilian shipping. Mistakes can put neutral ships at risk and create unintended escalation.

The sequence places the dataset's publication alongside a subsequent dispute over the success of a maritime strike:

• January 2026: The DMSD research appeared in the Journal of Radars, describing more than 2,000 paired images and nearly 20,000 annotated instances.
• February 2026: Iran claimed a successful strike on USS Abraham Lincoln near the Strait of Hormuz. Washington said the missiles or drones had not come close.

The dataset's practical role in training is established by its design and release; no date was given for its use in an operational weapons system. Its relevance lies in the recognition and tracking problems that naval surveillance and strike systems must solve.

The Institutions Behind the Project

The Naval Aeronautical University operates within China's military university system. It trains personnel for the People's Liberation Army Navy and develops technologies for naval aviation, making its participation directly relevant to aerial observation and maritime operations.

Harbin Engineering University contributes an institutional background in underwater acoustics and marine engineering. The Chinese Academy of Sciences conducts foundational research supporting national defence priorities, with its Institute of Computing Technology participating in this project.

The collaboration brings military and academic research into a single effort. Its institutional makeup points toward applications in unmanned aerial vehicles and precision strike systems, alongside the broader work of detecting and monitoring ships.

Publishing the research openly also fits a wider pattern in Chinese defence research: technical advances are made available to accelerate development while helping establish common methods for the domestic defence industry. DMSD gives researchers a common collection on which to train and assess their algorithms.

Why Release Military Relevant Data Publicly?

Public access allows civilian researchers as well as defence specialists to work on the same detection problems. Improvements to an algorithm developed in an academic setting can contribute to the wider body of methods available for maritime AI.

The release can serve several strategic purposes at once. It can speed the development of AI for naval technologies, encourage use of Chinese research methods as international benchmarks and place Chinese institutions near the centre of collaboration on maritime detection.

A shared dataset also establishes a technical starting point. Researchers can develop different models while using the same underlying collection, giving the data architecture a continuing role in how the problem is studied.

Potential applications include autonomous surveillance drones, precision missiles designed to strike ships and intelligent reconnaissance systems. These are relevant in environments where electronic interference disrupts conventional targeting or where GPS is unavailable.

The wider military interest is tied to strategies intended to keep opposing forces out of an area or restrict their freedom to operate within it. Such strategies depend in part on finding particular warships among other vessels and maintaining enough information to support an operational decision.

For deterrence and planning, the ability to recognise a moving target reliably matters alongside the ability to send a weapon toward it. DMSD concentrates on the training data behind that recognition step, including the difficult transition between a clear daytime view and a scene obscured by darkness or changing maritime conditions.

What the Release Changes for Researchers

DMSD gives the research community a freely available collection that its creators describe as larger than previous public ship detection datasets. More than 2,000 paired images and nearly 20,000 annotated instances put the emphasis on both the number of examples and the relationship between visible and thermal observations.

The research challenge remains concrete: locate the vessel, identify relevant differences, and keep following it as the ship, sensor and surroundings change. Precise labels help teach the first steps, while paired observations support work on using complementary sensor information.

The military significance comes from the same practical requirements. Surveillance platforms need to maintain contact, reconnaissance systems need useful identification, and weapons need accurate information during an engagement. The dataset offers a common foundation for developing the algorithms behind those tasks.

Key Points

• Chinese researchers released DMSD, containing more than 2,000 paired visible and infrared images and nearly 20,000 annotated instances.
• The team describes it as the world's largest publicly available ship detection dataset.
• The research appeared in the Journal of Radars in January 2026 and involved three military and academic institutions.
• The collection supports work on recognising vessels amid glare, darkness, weather changes and radar interference.
• Potential uses include surveillance drones, reconnaissance systems and precision missiles designed to strike ships.
• Iran claimed a February strike on USS Abraham Lincoln near the Strait of Hormuz; Washington said the missiles or drones had not come close.
• Public access gives researchers a shared resource for improving maritime AI and comparing detection methods.