IW Extraction From SAR Images Based on Generative Networks With Small Datasets

Internal waves (IWs), a significant phenomenon in various marine environments, play a crucial role in sustaining marine ecosystem balance. Accurate extraction of IWs information is essential for studying their properties. Most existing deep-learning-based internal wave extraction models rely heavily on large training datasets. However, large amounts of IWs synthetic aperture radar (SAR) images are difficult to access in practice. To address this issue, this article developed a generative adversarial network with multiscale downsampling and upsampling (GAN-MSDU) consisting of a series of adversarial networks at multiple scales. The model determines the scale through upsampling and downsampling procedures. Moreover, the adversarial network corresponding to each scale consists of a generator and a discriminator. The generator is designed to generate an IW as realistically as possible. The objective of the discriminator is to distinguish the generated image from the real data as much as possible. Through the design of multiscale architecture, this model successfully captures the characteristics of global and local IWs, and its adversarial training mode enhances the model’s representational ability via the generated data. The main contributions are as follows: first, to address the scarcity of IW image data, this study proposes a framework that requires only four images as the training set, providing a novel approach to the problem of data scarcity. Second, to address the elongated features of IW crests, this study proposes a novel downsampling method that preserves the crest features during the downsampling process. As a result, the overall mean accuracy, F1-score, mean intersection over union, and FWIoU of the GAN-MSDU model are 99.48%, 42.64%, 63.77%, and 99.30%, respectively. The comparison and quantitative evaluation with other methods for small data problems show that the GAN-MSDU method is efficient and robust in IW extraction.