A Change Detection Method Based on a Dual-Path Siamese Network With Boundary Feature Difference Modeling

Abstract

High-resolution remote sensing change detection remains challenging because pseudochanges, complex land-cover textures, and blurred object boundaries often degrade feature discrimination. This article presents the dual-path siamese multistage U-Net (DSM-UNet), a deep learning architecture that explicitly incorporates boundary feature difference modeling for accurate change detection. The framework employs a dual-branch encoder, in which a convolutional neural network pathway extracts fine-grained local details and a transformer pathway models long-range global dependencies. An improved global-local feature fusion mechanism is designed to integrate these complementary representations through spatial attention and adaptive weighting. To further suppress pseudochange noise and refine change boundaries, the differential feature multiscale fusion module extracts multiscale bitemporal differential features, while the feature guided attention module uses high-level semantic guidance to enhance detail-rich shallow features. Extensive evaluations on three public benchmark datasets, including LEVIR-CD, WHU-CD, and SYSU-CD, demonstrate that DSM-UNet achieves competitive or state-of-the-art performance. Ablation studies confirm the effectiveness of the proposed components, and complexity analysis indicates that DSM-UNet offers a favourable balance between accuracy and computational cost. The proposed architecture advances remote sensing change detection by jointly modeling global-local context, multiscale feature differences, and boundary-sensitive change cues.