NDWI-DeepLabv3+: High-Precision Extraction of Water Bodies from Remote Sensing Images

Abstract

How to efficiently and accurately extract water bodies from remote sensing images is the focus of scholars' research. Current research often does not make full use of the unique multi-band data of remote sensing images. This paper proposes an improved NDWI-DeepLabv3+ network to improve the accuracy of water body extraction, especially from urban remote sensing images. We improve the network from two main aspects: multi-scale input and multi-band data feature fusion. And for the critical parts of the network, we put forward a variety of feasible solutions to compare and select the best. In the end, we chose to convert the feature map calculated by NDWI into an input adapted to the neural network, and at the same time, develop a parallel convolution structure to fuse and extract the band data features. We verify the effectiveness of this method by comparing other multi-scale architecture networks in the same period. The NDWI-DeepLabV3+ network proposed in this paper can extract water from the L2A level data of Sentinel-2, which can slightly increase the computational consumption and obtain better performance. It provides new ideas for intelligently extracting hydrological information from remote sensing images.