Translation-classification loss for SAR image understanding with deep learning

SAR-to-optical translator networks are especially used to overcome the lack of optical images under cloudy conditions. Those translations being used for downstream tasks, they require the reconstruction of reliable patterns with respect to the underlying objects. In this paper, we propose a novel training strategy to account for land-cover complexity through a conjoint Translation-Classification Loss (TCL). The proposed loss evaluates the classifiability of translated images with a pre-trained land-cover classifier by assessing the reliability of its predictions and the relevance of its extracted hidden features. This new loss is applied to nine translators from the literature and to a tenth architecture introduced in the paper. Experiments show that applying the TCL not only improves the credibility of structures, patterns and textures but it also allows for better class discrimination and transitions while avoiding unreliable hallucinated artifacts produced by standard losses in adversarial approaches.