Image Analysis and Stereology · Published 2026-06-23 · DOI 10.5566/ias.3950
Rui Xu, Yihui Yang, Renzhong Mao, Yining Zhang, Yiteng Lin, Weiping Zhang
To address the issues of blurred building boundaries, small-object omission, and severe background interference in the semantic segmentation of high-resolution remote sensing imagery, this study proposes an improved method based on the Enhanced Vision Transformer Network (EViT). Specifically, this paper introduces a Grouped Cross-Cascaded Multi-Head Self-Attention (GCC-MSA) module to enhance feature diversity while maintaining linear complexity, and a Local-Global Feature Calibration (LGC) module to fuse CNN local details with Transformer global context. Coordinate Attention (CoAt) replaces conventional channel attention to strengthen channel-spatial feature representation. Additionally, Semantic-Guided Spatial Pyramid Pooling (SGSPP) and a GCC-MSA-guided Edge Perception (GEP) module reinforce multi-scale semantic perception and boundary extraction, while a Spatial Perception Gating Mechanism (SPGM) adaptively fuses dual-branch features. On the WHU Aerial, Massachusetts, and GF-7 Building Datasets, the model achieves Intersection-over-Union (IoU) scores of 92.33%, 77.81%, and 78.29%, respectively. These represent improvements of 0.57, 0.67, and 0.62 percentage points over the original EViT. The model demonstrates superior performance in small-building extraction, complex boundary segmentation, and background noise suppression, thereby providing a robust solution for precise surface object information extraction from high-resolution remote sensing imagery.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →
Xu, R., Yang, Y., Mao, R., et al. (2026). An Improved EViT Network for Semantic Segmentation of High-Resolution Remote Sensing Imagery. Image Analysis and Stereology. https://doi.org/10.5566/ias.3950