Multi-scale context extractor network for water-body extraction from high-resolution optical remotely sensed images

Abstract Water-body surveying and mapping is of great significance for water resources utilization, flood monitoring, and environmental protection. However, due to distribution diversities, shape and size variations, and complex scenarios of water-bodies, it is still challengeable to accurately and efficiently extract water-bodies from high-resolution remotely sensed images. In this paper, we propose a multi-scale context extractor network, termed as MSCENet, for delineating water-bodies from high-resolution optical remotely sensed images. The MSCENet mainly contains three key parts: a multi-scale feature encoder, a feature decoder, and a context feature extractor module. To address shape and size variations of water-bodies, the Res2Net is used in the feature encoder to extract rich multi-scale information of water-bodies. The context extractor module is composed of an assorted dilated convolution unit and a complex multi-kernel pooling unit, which further extracts multi-scale contextual information to generate high-level feature maps. The robustness and effectiveness of our MSCENet have been evaluated on two public datasets: LandCover.ai Data Set and DeepGlobe Data Set. Comparative experiments indicate the superiority and applicability of the MSCENet in water-body extraction.