Abstract
We present an algorithm for island wake segmentation from Synthetic Aperture Radar (SAR) imagery by incorporating a novel Chebyshev spatial correction branch into U-Net. Island wakes are important for sub-mesoscale air-sea turbulent energy transfer, yet are often excluded from forecasts due to a lack of real-time, automated retrievals. While high resolution SAR is well suited for detecting ocean surface roughness patterns associated with small scale ocean dynamics, island wake segmentation remains challenging. These difficulties arise from low signal-to-noise ratios and spatially varying additive and multiplicative biases, such as those resulting from incidence angle variations and uncorrected geometric distortions. Baseline U-Net is fully convolutional and therefore unable to model the spatially varying signals encountered in SAR imaging. To address these challenges, we propose a novel Chebyshev U-Net that incorporates a learned adaptive spatial bias correction branch that compensates for these effects. We evaluate our method against a baseline U-Net across ten Caribbean island locations using the European Space Agency's Sentinel-1 SAR data. Our results show that the Chebyshev U-Net qualitatively improves wake segmentation over the majority of the Caribbean islands and especially for images that exhibit strong limb darkening effects. Quantitative analysis reveals significantly improved IoU and F1-scores, although we observed a notable failure over Puerto Rico in cross validation. We further discuss strategies to address this failure case in future versions of the algorithm.