Masked Autoencoders for Enhancing SAR Sea Ice Image Segmentation

Synthetic aperture radar (SAR) imagery plays a crucial role in sea ice mapping. However, the performance of SAR image segmentation is often challenged by issues such as speckle noise and unstructured objects. In this study, a Vision Transformer (ViT) model is pretrained using a masked autoencoder (MAE), finetuned for semantic segmentation of ice types. Experimental results show that the MAE-based pre-training significantly enhances the ViT's ability to capture complex SAR image features, yielding a 5% improvement in F1-score over the UNet baseline, even with only 50% of the labeled data. The findings highlight the potential of self-supervised learning, particularly MAES, in addressing the challenges of limited labeled data in remote sensing tasks. Additionally, reconstruction performance indicates that the model effectively learns both the intrinsic features of SAR images, such as the incidence angle, and the complex structures of environmental objects, including ice, water, and land. This study demonstrates the potential of MAE-based pre-training as an effective strategy for SAR imagery analysis, offering more accuracy and efficiency in data-scare remote sensing applications.