Consistency-Driven Self-Supervised Feature Learning for Limited-Sample PolSAR Image Classification

Abstract

Polarimetric Synthetic Aperture Radar (PolSAR) images provide valuable information for identifying and monitoring different types of land cover. However, deep learning-based PolSAR classification generally requires large amounts of labelled training data, while obtaining reliable annotations is time-consuming and costly. Self-supervised learning offers an alternative by learning useful representations from unlabelled data before downstream classification. This study implements and systematically evaluates the Self-Supervised Learning with Multibranch Consistency (SSL-MBC) framework, originally introduced for few-shot PolSAR image classification. The framework exploits complementary PolSAR representations through multiple branches and employs consistency learning to obtain robust feature representations from unlabelled samples. The pretrained representation is subsequently transferred to a few-shot classification task using a limited number of labelled samples. The framework is evaluated on the Flevoland, San Francisco, and Oberpfaffenhofen PolSAR datasets. In addition to describing the architecture and training procedure in greater technical detail, this study presents dataset-specific experimental results and analyzes the contribution of multibranch representation learning under limited-label conditions. The revised experimental protocol aims to improve reproducibility and provide a clearer understanding of the effectiveness and limitations of SSL-MBC for few-shot PolSAR image classification. The results further support systematic assessment across datasets and label regimes and practical few-shot classification scenarios.