Internal Wave Signature Extraction from SAR Imagery Based On Deep Learning

In this study, we proposed an automatic internal wave signature extraction method for SAR (Synthetic Aperture Radar) imagery based on deep convolutional neural networks (DCNNs). For model training, we collected 116 labeled ENVISAT (Environmental Satellite) ASAR (Advanced SAR) images with clear internal wave signatures in the northern South China Sea. The specially tailored deep-learning-based extraction network leverages three modifications of U-Net to improve the accuracy and robustness. As a result, the overall mean Precision, Recall, and F1-score of the model are 84.95%, 84.71%, and 84.83%, respectively. The statistical results imply that our method can significantly increase the internal wave signature extraction accuracy from SAR images, even under complicated imaging conditions.