Water and vegetation segmentation of remote sensing images based on fusion of cross-stage boundary features and deep semantic features

Extraction of water and vegetation-covered areas from unmanned aerial vehicle remote sensing images is crucial for environmental monitoring, protection, and resource management. However, the irregular shapes and varying scales of these areas pose challenges for accurate boundary delineation and category discrimination using existing methods. To mitigate these problems, we design a multiscale feature fusion network. First, a boundary refinement module is designed in the network to solve the incomplete contour segmentation of the target region boundary. It refines the boundary contour by extracting the position association of the contour details of each stage and generates cross-stage boundary features. Second, to ensure the accuracy of classification, a semantic enhancement module is designed to enhance the representation of deep semantic features by capturing the neighborhood information of depth features. Finally, the multiscale feature fusion module is applied to fuse all features in the network to improve the segmentation performance. Experimental results demonstrate that the model attains notable performance with a mean intersection over union of 87.27%, an overall accuracy of 96.48%, and a mean accuracy of 91.92%. This method exhibits superior capability in enhancing boundary contour precision and accurately classifying water and vegetation-covered areas compared with other networks.