Decimeter-resolution PolSAR image segmentation using Segment Anything Model

Deep neural networks have dramatically changed semantic segmentation, outperforming previous strategies. After a surge of development over the last decade, neural networks are now available as off-the-shelf tools, such as the Segment Anything Model, which aims at performing single-shot segmentation. Applying SAM to Synthetic Aperture Radar images would be of advantage, since high resolution SAR images required for training neural networks remain a scarcely available resource. However, like the large majority of neural networks solutions, SAM was trained on optical images only, and is therefore not optimized for SAR images. In this paper we assess how SAM could be successfully applied to our fully polarimetric SAR images. We explore its performances on 35-centimeter resolution X-band images and 1.5-meter resolution L-band images, and how the choice of polarimetric decomposition affects the segmentation results.