Multiscale Difference Feature-Fusion Network for Change Detection With Hyperspectral Remote Sensing Images

Zhiyong Lv, Haoran Wang, Wei Li, Ming Zhang

Abstract: Land-cover change detection with hyperspectral remote sensing images (HyperCD) has become attractive in the applications of remote sensing images. Many existing studies have indicated that attention mechanisms play an important role in HyperCD. However, methods based on attention enhancement for HyperCD require further improvement. In this letter, we propose a novel multiscale difference feature-fusion network (MDFN) to improve the detection performance of HyperCD. First, a submodule named multiattention feature enhancement (MAFE) module was designed and embedded on each scale in the backbone of the proposed MDFN to capture subtle changes. Second, with the motivation of exploring the feature connection of a target on different scales, the attention feature maps from each scale were fused via a proposed novel cross-scale residual fusion module (CS-RFM). Finally, a softmax function was adopted to generate a binary change detection map based on the fused features. Experimental results based on comparison with five existing related works indicated that the proposed MDFN not only has some advantages in improving change detection performance with real hyperspectral remote sensing images (HRSIs) but also exhibits superiority in the requirement of training samples that are preferred in practical applications. For instance, using only 5% of the training samples, the average accuracy (AA) on the Farmland dataset is improved by 0.63%. The code will be available at https://github.com/ImgSciGroup/2024-MDFN.