MADF-Net: a lightweight remote sensing building change detection network based on multiscale attention and deformable feature enhancement

View at publisher Summarise with VIRAT
Authors: Jianfei Zhang, Weizhi Wang
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Building change detection from remote sensing imagery plays a vital role in urban planning, environmental monitoring, and related geospatial applications. Nevertheless, existing approaches often suffer from excessive computational complexity and limited feature extraction efficiency, which hinders their performance on high-resolution imagery and in complex change scenarios. To overcome these limitations, this study introduces a lightweight building change detection network, termed MADF-Net, designed to enhance detection accuracy through multiscale attention and deformable feature enhancement while maintaining high computational efficiency. Specifically, MADF-Net adopts LWGANet as a lightweight backbone to effectively reduce model complexity. The proposed hierarchical feature integration module facilitates cross-level interaction between shallow structural details and deep semantic representations; the deformable change extraction module leverages deformable convolutions with learnable offsets to dynamically align bi-temporal features, thereby mitigating errors induced by geometric distortions; and the multiscale attention refinement module—comprising a channel-spatial attention module and a dual attention refinement module—adaptively enhances multiscale targets and fine-grained change regions. Experimental results indicate that, compared with existing mainstream methods, MADF-Net demonstrates advantages in terms of computational cost (2.30 G FLOPs) and parameter count (2.68 M), while achieving F1 scores of 91.51% and 94.60% on the LEVIR-CD and WHU-CD public datasets, respectively. These results verify the capability of the proposed method to effectively balance detection accuracy and computational efficiency.

Fields of Study & Subjects

Remote-Sensing Image Classification Automated Road and Building Extraction Advanced Neural Network Applications

Publication details

Journal / source
Journal of Applied Remote Sensing
Volume
20
Issue
01
Pages
1–20
Publication date
18 March 2026
Language
EN