Sublook2Sublook: A Self-Supervised Speckle Filtering Framework for Single SAR Images

Speckle reduction is a preprocessing for synthetic aperture radar (SAR) image interpretation and application. With the advances of convolutional neural network (CNN) models, excellent speckle filters have been continuously developed. However, supervised learning-based models suffer from a generalization deficiency due to the lack of clean SAR images. Additionally, other self-supervised learning (SSL)-based methods rely on multiple independent SAR images of the same scene to generate the filtered SAR image. In practice, these additional auxiliary datasets are not always available, which limits the application of these methods. To fulfill this gap, a novel self-supervised framework named Sublook2Sublook is proposed for single SAR images speckle filtering. The main contribution of this work lies in that a new theorem is founded which guarantee the cost function defined on the paired sublook SAR images is statistically equivalent to the supervised counterpart based on the speckled-clean SAR image pairs. Thereby, the sublook SAR images can be alternatively used for model training instead of the clean SAR images or additional auxiliary datasets. From this fashion, a complete self-supervised speckle filter is developed. First, sublook decomposition is performed in both azimuth and range directions. Then, optimal paired sublook images are selected based on a criteria of minimum L1 norm distance. Finally, the established self-supervised speckle filter can be trained with the paired sublook images. Extensive experimental studies are conducted with various SAR datasets in terms of different frequency bands and spatial resolutions from the Radarsat-2, COSMO-SkyMed, and ALOS-2 SAR satellites. Comparisons studies with four state-of-the-art despeckling methods confirm the superiority of the proposed method. The results demonstrate that the proposed Sublook2Sublook framework can better smooth speckles in homogeneous areas while well preserve image details.