Multimodal self-supervised learning for semantic analysis of PolSAR imagery

The automatic prediction of semantic maps, e.g. for land use/cover, is often addressed via supervised learning approaches that rely on large training datasets. Obtaining the required reference annotations for a large amount of data is tedious and time consuming work, in particular for sensors such as synthetic aperture radar. Self-supervised learning aims to mitigate this issue by leveraging the large amount of usually available unlabelled data by constructing a pretext task that does not require annotations. We evaluate DINO, a recent self-supervised learning approach, and show that it outperforms segmentation models that have been initialized randomly or by weights obtained from pretraining on ImageNet. Furthermore, we show that the inclusion of multimodal data during the pretraining phase improves results significantly.