Large-scale mapping of water bodies across sensors using unsupervised deep learning

Abstract

Rapid and accurate monitoring of surface water is critical for water resource management, environmental protection, sustainable urban development, among other issues. Landsat and Sentinel data are publicly available optical data with high spatial and temporal resolution, providing the possibility for large-scale surface water mapping. However, traditional threshold-based or supervised classification-based surface water mapping methods often require adjusting thresholds or training samples for different areas or different sensors, which may hinder the generalization performance of the method in large-scale water body mapping. To address these difficulties, we propose an unsupervised cross-sensor deep learning water bodies mapping framework (UUCP) for unlabeled large-scale optical remote sensing images. The UUCP framework adopts an unsupervised multi-segment thresholding strategy to achieve the transition from label-free learning to noisy label learning. It learns robust multi-scale features of water bodies by the developed channel attention multi-scale surface water extraction network and training strategies under noisy labels. The proposed algorithm's effectiveness was evaluated using Sentinel-2 and Landsat-8 images from Guangzhou and Wuhan in China, and nine regions in France. The results show that our proposed method performs well in the overall performance of water extraction and is applicable to different sensors, with Kappa values reaching an average of 0.8859 and 0.8084 on Sentinel-2 and Landsat-8, respectively. More importantly, in cross-sensor experiments (the model trained on Landsat-8 data directly predicts Sentinel-2 dataset), the UUCP algorithm has excellent performance and is superior to other traditional water extraction algorithms. Overall, UUCP has excellent generalization ability and provides a new perspective for large-scale surface water mapping.