SAR-W-MixMAE: Polarization-Aware Self-Supervised Pretraining for Masked Autoencoders on SAR Data

Self-supervised pretraining has emerged as a powerful approach for learning transferable representations from large-scale unlabeled data, significantly reducing reliance on taskspecific labeled datasets. Although masked autoencoders (MAEs) have shown considerable success in optical remote sensing, such as RGB and multispectral imagery, their application to synthetic aperture radar (SAR) data remains underexplored due to its unique imaging characteristics, including speckle content and intensity variability. In this work, we investigate the effectiveness of masked autoencoders for SAR pretraining, specifically applying MixMAE [1] to Sentinel-1 SAR imagery. We introduce SARW- MixMAE, a domain-aware self-supervised learning approach that incorporates a SAR-specific pixel-wise weighting strategy into the reconstruction loss, mitigating the effects of speckle content and high-intensity backscatter variations. Experimental results demonstrate that SAR-W-MixMAE consistently improves baseline models in multilabel SAR image classification and flood detection tasks, extending the state-of-the-art performance on the popular BigEarthNet dataset. Extensive ablation studies reveal that pretraining duration and fine-tuning dataset size significantly impact downstream performance. In particular, early stopping during pretraining can yield optimal downstream task accuracy, challenging the assumption that prolonged pretraining enhances results. These insights contribute to the development of foundation models tailored for SAR imagery and provide practical guidelines for optimizing pretraining strategies in remote sensing applications.