EMDNet: edge-aware multi-scale difference network for remote sensing image change detection
Abstract
In high-resolution remote sensing, change detection aims to identify land-cover transitions between bi-temporal images, yet blurred boundaries, illumination-induced pseudo-changes, and sparse change signals under severe changed-unchanged imbalance remain major obstacles. To address these issues, an edge-aware multi-scale difference network (EMDNet) is proposed with progressive stages for encoding, difference reasoning, and decoding. An edge-aware central-difference stem (ECF-ResNet18) is designed to retain high-frequency boundary cues before downsampling, helping alleviate early edge attenuation that blurs boundaries and misses small objects. A multi-scale spatiotemporal transformer module (MSST-Module) is introduced to jointly reason over connection, residual, and appearance streams for spectral-spatial-temporal difference modeling, helping separate true material changes from illumination-driven pseudo-changes. An adaptive region-aware progressive decoder (ARP-Decoder) is designed with complementary dual gating to recover sparse change responses when unchanged pixels dominate under severe class imbalance. Together, the three stages link edge-preserving encoding, difference reasoning, and region-aware decoding so that each prepares the representations required by the next. Compared with recent CNN-, transformer-, and Mamba-based methods on LEVIR-CD, SYSU-CD, and WHU-CD, EMDNet achieves F1 scores of 91.09%, 83.01%, and 90.70% and OA of 99.10%, 92.37%, and 99.31%, with a practical inference cost of 13.14 ms per 256×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document}256 patch. These results demonstrate that EMDNet improves boundary fidelity over recent strong baselines. Pseudo-change control and sparse-change recovery are also strengthened, while inference cost remains practical.