An End-to-End Point-Supervised Method for Semantic Segmentation of High-Resolution Remote Sensing Images

Deep learning networks have become a new paradigm for semantic segmentation of remote sensing images. However, their great achievements heavily rely on a large number of high-quality labels, which costs a lot in acquisition. To cope with this issue, this paper introduced point labels, which provided only the categories of a small number of pixels. To fully exploit point labels for semantic segmentation, this paper proposed a simple but effective end-to-end point-supervised method, which combined a dense energy loss for learning the spatial relations of unlabeled pixels and a partial cross-entropy loss for learning labeled pixels. The Vaihingen and Zurich summer datasets were tested for the experiments. Results show that the proposed method can improve the mean F1 values by 1.57%-5.12%, compared to the baseline, and it can also retain well the boundary information of small objects.