Masked auto-encoding and scatter-decoupling transformer for polarimetric SAR image classification

The pixel level annotation of polarimetric SAR (PolSAR) image is quite difficult and requires a significant amount of manpower. Deep learning based PolSAR image classification generally faces the challenge of scarce labeled data. To address the above issue, we propose a self-supervised learning model based on masked auto-encoding and scatter-decoupling transformer (MAST) for PolSAR image classification, which aims to fully utilize a large number of unlabeled data. Combined with PolSAR scattering characteristics, an effective pre-training auxiliary task is designed to constrain the model in order to learn spatial information and global scattering representation from SAR images. In the fine-tuning stage, a scattering embedding module is applied to strengthen the representation of global semantic information with specific scattering characteristics. In addition, a supervised contrastive loss is introduced to improve the robustness of the classifier. Sufficient experiments are conducted on three public PolSAR datasets, and the results demonstrate the effectiveness of the proposed method.