A water edge extraction method based on deep learning for remote sensing images

The paper proposes a pixel-level sea-land semantic segmentation network A-Unet based on the attention mechanism to extract water edges. The network refines the A-Unet classification results through conditional random fields to realize the pixel-level semantics of remote sensing image water edges segmentation. Using the historical remote sensing images of Tianjin coastal area and Weihai area as data source, the paper extracts the coastline of Tianjin area of the past ten years and anaylses its change trendency qualitatively and quantitatively. Moreover, the natural waterline of Weihai area is employed to verify the accuracy of proposed model. Experiments show that compared with other waterfront segmentation methods, the proposed network can obtain more refined results, which can provide a better decision-making and reference basis for the rational development of urban marine resources and the protection of marine ecological environment.