Semantic Segmentation of Water Body in High-Resolution Remote Sensing Images Based on DeepLabV3+

Abstract

Using image semantic segmentation technology to realize remote sensing water recognition is the frontier direction of current remote sensing water recognition research. In order to improve the problems of low segmentation accuracy and incomplete extraction when the current semantic segmentation method of remote sensing image water body involves complex background and small area water body and other scenes, this paper takes remote sensing water body image as the object of study and designs a water body segmentation method based on DeepLabV3+, which improves and designs the original structure prone to grid effect. The part of ASPP that causes information loss; Combining with the characteristics of binary classification of water extraction task, an attentional feature fusion module AFFM was designed and constructed. The improved ASPP module and AFFM module were applied to DeepLabV3+ network, which effectively improved the accuracy of semantic water segmentation in remote sensing images.