Joint compression and despeckling by SAR representation learning

Abstract

Synthetic Aperture Radar (SAR) imagery is a powerful and widely used tool in a variety of remote sensing applications. The increasing number of SAR sensors makes it challenging to process and store such a large amount of data. In addition, as the flexibility and processing power of on-board electronics increases, the challenge of effectively transmitting large images to the ground becomes more tangible and pressing. In this paper, we present a method that uses self-supervised despeckling to learn a SAR image representation that is then used to perform image compression. The intuition that despeckling will additionally improve the compression task is based on the fact that the image representation used for despeckling forms an image prior that preserves the main image features while suppressing the spatially correlated noise component. The same learned image representation, which can already be seen as the output of a data reduction task, is further compressed in a lossless manner. While the two tasks can be solved separately, we propose to simultaneously training our model for despeckling and compression in a self-supervised and multi-objective fashion. The proposed network architecture avoids the use of skip connections by ensuring that the encoder and decoder share only the features generated at the lowest network level, namely the bridge, which is then further transformed into a bitstream. This differs from the usual network architectures used for despeckling, such as the commonly used Deep Residual U-Net. In this way, our network design allows compression and reconstruction to be performed at two different times and locations. The proposed method is trained and tested on real data from the TerraSAR-X sensor (downloaded from https://earth.esa.int/eogateway/catalog/terrasar-x-esa-archive). The experiments show that joint optimization can achieve performance beyond the state-of-the-art for both despeckling and compression, represented here by the MERLIN and JPEG2000 algorithms, respectively. Furthermore, our method has been successfully tested against the cascade of these despeckling and compression algorithms, showing a better spatial and radiometric resolution, while achieving a better compression rate, e.g. a Peak Signal to Noise Ratio (PSNR) always higher than the comparison methods for any achieved bits-per-pixel (BPP) and specifically a PSNR gain of more than 2 dB by a compression rate of 0.7 BPP.