Extracting Lakes and Reservoirs From GF-1 Satellite Imagery Over China Using Improved U-Net

Abstract

Lakes and reservoirs (LaR) are important parts of water resources and their rapid and accurate monitoring is an essential guarantee for maintaining ecological health and social development. The existing waterbody extraction methods are mostly targeted at local water bodies, with little attention on the national scale. In this letter, an improved U-Net method is proposed for LaR extraction from GF-1 satellite imagery. First, 21 scenes of GF-1 images are evenly selected across China, and the training set and validation set are produced by image processing, cropping, and augmentation. Second, a deep learning network is constructed by modifying the U-Net, deepening the network and introducing multiple skip connections, which is suitable for extracting LaR China-wide. Experiments on the GF-1 imagery demonstrate that the superiority of the improved U-Net when compared with other deep learning methods (U-Net, UNet++, FastFCN, DeepLabv3+) and traditional methods [the normalized difference water index (NDWI), maximum likelihood method (MLM)]. In addition, 20 LaR are selected for further evaluation of the model, and all of them achieve good extraction results, showing excellent generalization of the model.