Self-supervised learning for segmentation of polarimetric SAR imagery

Kurzfassung

Semantic segmentation of Polarimetric Synthetic Aperture Radar (PolSAR) imagery is crucial for environmental monitoring, but the segmentation accuracy is often limited by the scarcity of annotated data. Acquiring reliable ground-truth labels for PolSAR is inherently challenging, often requiring complex and logistically demanding field campaigns. This study investigates self-supervised learning (SSL) to mitigate this limitation using the Pol-InSAR-Island benchmark dataset. We propose a framework that employs a Masked Autoencoder (MAE) to learn robust feature representations from unlabeled PolSAR data, which are subsequently fine-tuned within a U-Net architecture for semantic segmentation. We conducted a comprehensive ablation study to compare the SSL-pretrained model against a fully supervised baseline. This analysis systematically evaluated how different data representations, normalization strategies, and data augmentation techniques affect model performance. The results demonstrate a substantial performance gain from SSL pretraining, boosting the mean Intersection over Union (IoU) from 21.71\% (supervised baseline) to 36.93\%. Furthermore, the pretraining enhanced training stability, halving the coefficient of variation (CV) across runs from 1.22 to 0.66. Our analysis confirmed that an extended log-ratio data representation combined with a trimmed standardization and clipping normalization strategy yielded the best performance. While data augmentation techniques like CutMix offered moderate improvements, the contribution from SSL pretraining was markedly more impactful, especially for segmenting classes with complex, heterogeneous boundaries. These findings establish SSL as a highly effective strategy for PolSAR semantic segmentation, demonstrating that powerful, transferable features learned from unlabeled data can enable high-accuracy classification even with a severely limited number of labeled samples.