Self-Supervised Learning With Multibranch Consistency for Few-Shot PolSAR Image Classification

Abstract: Polarimetric Synthetic Aperture Radar (PolSAR) images provide valuable information for land-cover classification and can capture surface characteristics even under challenging weather and lighting conditions. However, training deep learning models for PolSAR image classification generally requires a large number of labelled samples, which are often difficult and expensive to obtain. This limitation becomes more significant in few-shot classification, where only a small number of labelled samples are available for each class. This paper presents a Self-Supervised Learning with Multibranch Consistency (SSL-MBC) approach for few-shot PolSAR image classification. The proposed method first learns useful image representations from largely unlabelled PolSAR data using a self-supervised learning strategy. A shared feature encoder is combined with multiple convolutional branches using different kernel sizes to capture spatial and contextual information at different scales. The consistency between the representations generated from different views of the same PolSAR image is used to improve the quality and stability of the learned features. After pretraining, the learned representations are transferred to a classification network and fine-tuned using only a small number of labelled samples. The proposed approach is evaluated on PolSAR land-cover classes such as water, forest, urban, and cropland. Experimental results show that the method can achieve improved classification performance compared with conventional supervised deep learning models under limited-label conditions. The results demonstrate that combining self-supervised representation learning with multibranch consistency can reduce the dependence on labelled data while maintaining reliable classification performance. The proposed framework therefore provides a practical approach for PolSAR image classification in scenarios where obtaining large labelled datasets is difficult.

Keywords: Keywords: PolSAR Image Classification, Self-Supervised Learning, Few-Shot Learning, Multibranch Consistency, Deep Learning, Land-Cover Classification, Representation Learning, Synthetic Aperture Radar (SAR).