Multiscale Residual Alignment Transformer for Remote Sensing Image Change Detection
- 27 May 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 22 (1545598X) , 1-5
Abstract
Deep learning (DL) methods have shown great potential for remote sensing image change detection recently, but still suffer from several limitations. Within the identical semantic concept, significant but irrelevant changes in surface texture, color and spatial shifting of building objects resulted from variations in imaging physical factors, causes feature inconsistency of building objects in bitemporal sences. Conventional DL methods lack the capability to effectively distinguish real changes from irrelevant changes, leading to some false detections. This letter propose a novel framework, the multi-scale residual alignment transformer (AlignFormer), to mitigate the above issues. Specifically, inspired by deformable attention mechanism, we firstly design an adaptive feature alignment module (AFAM) to suppress the inconsistency of feature pairs, where the regions of building can be adaptively focused on and the spatial-temporal dependencies of relevant building objects in feature pairs effectively captured, via scheme of flexibly sampling Keys/Values for each given Query. Besides, we utilize an extremely tiny Swin Transformer as the backbone of differencing-based framework for obtaining hierarchical features. Moreover, inspired by residual learning strategy, three AFAMs are integrated into the framework to form the multi-scale residual architecture for the coarse-to-fine alignment of the paired features. Experimental results confirm the superiority of our proposed method over several state-of-the-art algorithms. Our code will be released at https://github.com/lilei-aircas/AlignFormer_CD.Keywords
Funding Information
- Key Scientific Research Projects Deployed by Chinese Academy of Sciences (KGFZD-145-23-24)
This publication has 29 references indexed in Scilit:
- Edge-Guided Feature Dense Fusion Network for Remote Sensing Image Change Detection2023
- A Transformer-Based Siamese Network for Change Detection2022
- Vision Transformer with Deformable Attention2022
- Swin Transformer: Hierarchical Vision Transformer using Shifted Windows2021
- DASNet: Dual Attentive Fully Convolutional Siamese Networks for Change Detection in High-Resolution Satellite ImagesIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021
- A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing imagesISPRS Journal of Photogrammetry and Remote Sensing, 2020
- Fully Convolutional Siamese Networks for Change Detection2018
- Building Change Detection From Multitemporal High-Resolution Remotely Sensed Images Based on a Morphological Building IndexIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2013
- A change detection model based on neighborhood correlation image analysis and decision tree classificationRemote Sensing of Environment, 2005
- Change vector analysis: A technique for the multispectral monitoring of land cover and conditionInternational Journal of Remote Sensing, 1998