CSFMamba: A FrFT-Enhanced Cross-State Fusion Network for GNSS-R Sea Surface Height Retrieval in Complex Coastal Waters

The ocean is never still. Its surface rises and falls with the pull of the moon, the push of the wind, and the flow of rivers, creating a constantly shifting landscape that is difficult to map from space. For decades, scientists have used satellites to measure the height of the sea, a task critical for understanding climate change, predicting storms, and navigating ships. However, looking at the open ocean is one thing; looking at the coast is another. Near the land, the water is shallow, the tides are wild, and the signal from a satellite can bounce off the beach or a nearby island before it hits the water, confusing the instruments. This "coastal problem" has long been a blind spot in our ability to monitor the world's seas.

To solve this, researchers are turning to a clever trick using the same signals that guide our smartphones. Satellites orbiting the Earth constantly beam down navigation signals. When these signals hit the ocean, they bounce off the surface and return to a receiver, sometimes on another satellite. By measuring how long the signal takes to travel and how its shape changes after bouncing, scientists can calculate the height of the water. This method, known as GNSS-R, works beautifully over the deep, open ocean. But in complex coastal waters, the signals get messy. The reflection is distorted by the roughness of the waves, the angle of the sun, and the interference from nearby land. Traditional methods, which rely on fixed physical formulas to correct these errors, often struggle in these chaotic environments, leading to inaccurate measurements.

A team of researchers from Guilin University of Technology in China has developed a new approach to tackle this specific challenge, focusing on the Beibu Gulf, a semi-enclosed bay in the South China Sea known for its complex tides and monsoon-driven weather. They created a new computer system called CSFMamba, which uses a type of artificial intelligence designed to learn from complex patterns rather than just following rigid rules. Instead of trying to force the messy coastal data into a simple equation, this system learns to recognize the subtle relationships between the shape of the reflected signal and the actual height of the water, even when the conditions are difficult.

The core of their innovation lies in how the system processes the data. The raw information comes in the form of a delay-Doppler map, which is essentially a detailed snapshot of the signal's energy as it bounces off the sea. In the open ocean, this snapshot is usually clear, but near the coast, it is often blurred and noisy. The researchers first applied a mathematical technique called the fractional Fourier transform to sharpen these blurry snapshots, making the hidden patterns of the water's surface more visible. They then fed this enhanced data into a dual-branched neural network. One branch of the network acts like a visual expert, studying the shape and texture of the signal map to understand the local sea state. The other branch acts like a context expert, analyzing a wide range of numbers including wind speed, rainfall, the angle of the satellite, and the time of day.

What makes this system unique is how these two branches talk to each other. In many previous models, different types of data were simply stacked together or added up, which often failed to capture how the environment and the signal shape influence one another. The CSFMamba system uses a mechanism called cross-state fusion, which allows the visual and numerical branches to dynamically adjust their focus based on the current situation. If the wind is high, the system learns to weigh the wind data more heavily; if the signal is distorted by land, it learns to rely more on the geometric position of the satellite. This dynamic interaction allows the model to adapt to the specific, chaotic conditions of the Beibu Gulf in a way that static models cannot.

The researchers tested their system using data from 2021 and 2022, comparing its predictions against established global sea level models. The results were striking. Over the four-month test period, the system achieved an average error of just 0.53 meters, a significant improvement over existing methods. To put this in perspective, if you were trying to measure the height of a person standing on a boat, this level of accuracy means you could tell the difference between a child and an adult with high confidence, even while the boat was rocking. The system maintained this high level of accuracy across different seasons, from the calm transition periods to the turbulent summer monsoons, proving that it could handle the changing weather patterns of the region.

To ensure the system was truly learning the right things and not just memorizing the data, the researchers performed a series of tests where they removed specific parts of the model. They found that the most critical component was the fusion mechanism that allowed the two branches to interact. When this part was removed, the error rate more than doubled, dropping the system's performance to a level where it could no longer be trusted for precise measurements. This confirmed that the ability to blend the visual signal patterns with the environmental context was the key to success. Furthermore, the system's internal analysis showed that it correctly identified the most important factors, such as the geographic location and the wind speed, aligning with our physical understanding of how the ocean behaves.

The study also compared the new system against other advanced artificial intelligence models, including those based on traditional deep learning and newer attention-based architectures. The CSFMamba system consistently outperformed these competitors, reducing the average error by roughly a quarter compared to the best existing alternatives. This suggests that for complex coastal environments, the specific architecture of this new model is better suited to the task than previous approaches. The researchers noted that while the system works well in the main body of the gulf, it still faces challenges in the most difficult areas, such as right next to ports or in narrow river estuaries where land interference is extreme. These areas remain difficult because the signals are so heavily distorted by the nearby geography that even the most advanced learning models struggle to find a clear pattern.

Despite these remaining challenges, the work represents a significant step forward in our ability to monitor the world's coastlines. By successfully applying this new method to the Beibu Gulf, the researchers have demonstrated that it is possible to retrieve accurate sea level data from space even in the most turbulent and complex nearshore waters. This capability opens the door to better monitoring of coastal flooding, more accurate storm surge predictions, and a deeper understanding of how rising sea levels are affecting the world's most populated regions. The success of the CSFMamba system suggests that the future of coastal oceanography lies not in simpler formulas, but in intelligent systems that can learn to navigate the messy, beautiful complexity of the real world.