Hybrid Multi- Expert Deep Learning Framework for Operational SAR Maritime Surveillance and Dark Vessel Identification

Imagine the ocean as a giant, bustling highway where ships are the cars. For decades, we've tracked these cars using a system called AIS, which is like a digital license plate that ships voluntarily broadcast to say, "Here I am!" But just like a driver might turn off their headlights to sneak past a toll booth, some ships turn off their AIS signals to hide. These are called "dark vessels," and they can be used for smuggling, illegal fishing, or other shady business. To catch them, we need eyes that can see through the fog, rain, and pitch-black night. Enter Synthetic Aperture Radar, or SAR. Think of SAR as a super-powered flashlight that doesn't use light, but uses invisible microwave beams. It can bounce off ships and create a picture even when it's storming or the moon is hidden. However, looking at these radar pictures is tricky. They are often grainy, full of static (like an old TV), and confusing, especially near crowded ports where ships are packed together like sardines. Scientists have tried to use computers to find these ships, but most computer programs are like generalists: they are okay at spotting a ship in the open ocean, but they get confused in a busy harbor or a rocky coast.

This is where a new study by researchers from Amity University steps in. They propose a "Hybrid Multi-Expert Deep Learning Framework," which is a fancy way of saying they built a team of specialized computer detectives instead of relying on just one. Imagine a police station where you don't send every detective to every crime scene. Instead, you have a smart dispatcher who looks at the scene first. If it's a quiet, open ocean, the dispatcher sends in "Expert A," a fast and efficient detective. If it's a choppy coastline, they send in "Expert B," who is a bit more careful. If it's a super crowded port with ships overlapping each other, they send in "Expert C," the most advanced detective with special tools to untangle the mess. The researchers tested this team using real satellite images and found that this "smart dispatcher" system works incredibly well. They showed that by matching what the radar sees with the "digital license plates" (AIS) that ships do broadcast, they can spot the ones that are hiding. In a test over the busy Singapore Strait, the system found over 1,000 ships and identified 96 of them as "dark vessels" that were trying to stay invisible.

The Problem with the "One-Size-Fits-All" Detective

For a long time, scientists have tried to use a single computer program to find ships in radar images. It's a bit like trying to use a single pair of glasses to read a book, drive a car, and look at the stars all at once. It might work okay for one thing, but it's not great for the others. In the open ocean, ships are far apart, so a simple, fast detector works fine. But near the coast, the water is rough, and the radar gets confused by the waves. In a dense port, ships are so close together they look like one giant blob. Traditional methods, like the "Constant False Alarm Rate" (CFAR) technique, often get overwhelmed by this background noise, ringing the alarm for every wave and missing the actual ships.

The researchers argue that we need a system that adapts to the environment. They suggest that instead of forcing one model to do everything, we should let the computer decide which "expert" to use based on what it sees.

The Smart Dispatcher: Routing the Scene

The first step in this new framework is the "Intelligent Scene Routing." Think of this as the dispatcher at the front desk. When a new radar image arrives, the system doesn't just start looking for ships immediately. First, it uses a modified version of a neural network called ResNet-50 to look at the picture and ask, "Where are we?"

The system classifies the scene into three types:

1. Offshore: The open sea, where it's relatively quiet and clear.
2. Inshore: The coast, where the water is choppier and the background is messier.
3. Dense Port: A crowded harbor where ships are packed tight.

The researchers trained this dispatcher to be very good at this job. In their tests, it correctly identified the type of environment 98.8% of the time. This is crucial because it ensures the right tool is used for the right job. If the dispatcher sends the wrong expert, the system might miss a ship or get confused by a wave.

The Team of Specialized Detectives

Once the dispatcher knows the location, it hands the image to the specific expert designed for that environment:

• Expert A (Offshore): For the open ocean, the system uses YOLOv8n. This is a lightweight, fast model. Since the open sea is less cluttered, this expert doesn't need to be overly complicated; it just needs to be quick and accurate. The researchers found this expert achieved a detection accuracy (called mean Average Precision or mAP) of 97%.
• Expert B (Inshore): For the coast, the system switches to YOLOv8s, which is slightly larger and more careful. It also uses a special trick called "CFAR-based adaptive filtering." Imagine this as a detective who adjusts their sensitivity based on the noise level. If the waves are loud, the detective knows to ignore the small ripples and focus on the big shapes. This expert achieved an mAP of about 86%.
• Expert C (Dense Port): This is the heavy lifter. In crowded ports, ships overlap, making them hard to separate. This expert uses a combination of SwinIR (a tool that cleans up the grainy radar image, making it sharper) and Mask R-CNN (a model that can draw a precise outline around each individual ship, even if they are touching). This expert was so good at finding overlapping ships that it achieved a "recall" of nearly 1.0, meaning it found almost every single ship that was there, missing very few.

Catching the Hiders: The AIS Fusion

Finding the ships is only half the battle. The real goal is to find the dark vessels—the ones that are hiding. To do this, the system uses a clever trick called "AIS Data Fusion."

Every ship that plays by the rules broadcasts its location via AIS. The system takes the list of ships it found with the radar and compares it to the list of ships broadcasting their location. It uses a mathematical formula (the Haversine distance) to see if a radar ship is close enough to an AIS ship to be the same one. They set a rule: if a radar ship is within 500 meters of an AIS ship, they are a match.

If the radar sees a ship, but there is no AIS signal nearby, the system flags it as a Dark Vessel. It's like seeing a car driving without headlights; you know it's there, but it's not telling you who it is.

Tracking the Mystery Ships

Once a dark vessel is spotted, the system doesn't just take a snapshot; it starts tracking. It uses a Kalman Filter, which is a mathematical method for predicting where something will go next based on where it has been. Think of it as a game of "hot potato" where the system predicts the ship's path. If a ship suddenly changes direction or stops moving in a weird way, the system can spot this anomaly. This helps authorities understand not just where the ship is, but what it might be doing.

The Results: Putting It to the Test

The researchers didn't just build this in a computer simulation; they tested it with real data. They used images from the Sentinel-1 satellite and data from the Singapore Strait, one of the busiest shipping lanes in the world.

Here is what they found:

• The system analyzed 1,038 ship detections.
• It identified 96 of them as dark vessels.
• This resulted in a Dark Vessel Detection Rate (DVDR) of 9.24%.

This means that in this specific test, about 1 in every 10 ships detected by the radar was not broadcasting its location. The researchers note that this aligns with previous studies, suggesting that a significant portion of maritime traffic is operating in the shadows.

Why This Matters

The main takeaway from this paper is that a "one-size-fits-all" approach to finding ships isn't good enough. By creating a system that acts like a team of specialists—each with their own superpowers for different environments—the researchers have built a more robust and accurate way to monitor the seas.

The study suggests that this hybrid framework is a strong step forward for operational maritime surveillance. It moves beyond simple simulations and shows that it can work with real satellite data and live traffic information. While the system isn't perfect (no system is), it offers a promising way to catch the "dark vessels" that try to slip through the cracks, helping to keep global trade routes safer and more secure. The researchers plan to keep improving the system, perhaps by combining radar with other types of sensors in the future, but for now, they have shown that a smart, multi-expert team is the best way to solve the puzzle of the hidden ocean.