Semi-supervised segmentation of individual buildings from SAR imagery

Buildings are essential geo-targets that contribute to the monitoring of urban development. Synthetic aperture radar (SAR) provides excellent opportunities for building segmentation as it is insensitive to sunlight illumination and weather conditions. Nevertheless, the majority of existing approaches that exploit convolutional neural networks (CNNs), need to collect an enormous quantity of annotations for network training. Therefore, we propose an innovative semi-supervised method for individual building segmentation from SAR imagery. Our approach has three modules: a weights-shared encoder, a main decoder as well as an auxiliary decoder. For unlabeled samples, given the perturbation added to the encoder's output, we enforce the consistency between the feature and output of the auxiliary decoder and those of the main decoder. This allows for the use of abundant unlabeled samples to make up for a lack of supervisory information. The experiments are carried out on a SAR dataset that is collected from the city of Berlin, Germany. Quantitative and qualitative results suggest that our approach is superior to other competitors.