Learning Deep Models from Weak Labels for Water Surface Segmentation in Sar Images

Today, one of the biggest challenges faced in the intersection of the Deep Learning (DL) and synthetic aperture RADAR (SAR) domains is the scarcity of precisely annotated datasets suitable for properly training a supervised algorithm. This paper shows that it is possible to successfully exploit weak-labeled data instead of relying on manually annotated labels. In particular, we show how it is possible to train, with state-of-the-art performance, a deep model for the segmentation of water surfaces in SAR images from a weak-labeled dataset. Finally, we present examples of applications of the learned model to the segmentation of inland water bodies and floods.