A Novel Transformer-Enhanced Framework for Multi-scale Feature Fusion in Remote Sensing Change Detection

Abstract

With the rapid advancement of remote sensing technology, high-resolution satellite imagery has become increasingly accessible for monitoring urban expansion and surface changes. However, change detection in remote sensing images faces several significant challenges, including inconsistent image registration, inadequate handling of multi-scale features, class imbalance in training samples, and the occurrence of pseudo-changes that appear as modifications but represent no actual surface alterations. To address these limitations, this paper presents MA-CDNeXt, a novel deep learning model that enhances the Transformer architecture for improved change detection in high-resolution remote sensing imagery. The proposed model incorporates three key innovations: (1) an enhanced feature encoder that better captures spatial-temporal information, (2) integration of MAFGNet's multi-scale attention fusion mechanism within CDNeXt's spatial-temporal module to effectively process features at different scales, and (3) an optimized feature fusion decoder equipped with a dual-branch graph convolution module for more accurate change localization. To mitigate the class imbalance problem commonly encountered in change detection datasets, we employ a combined loss function incorporating cross-entropy loss, Dice loss, and focal loss. Experimental evaluation on the LEVIR-CD+ dataset demonstrates the effectiveness of our approach, achieving an overall accuracy of 84.09% and a recall rate of 78.39%. The results indicate that MA-CDNeXt successfully addresses the key challenges in remote sensing image change detection and provides a robust solution for monitoring surface changes in high-resolution imagery.