Cascaded attention-induced difference representation learning for multispectral change detection

Abstract

Change Detection (CD) is the process of recognizing and quantitatively evaluating changes in surface objects at the exact location but at different times from remote sensing images that may be caused by extreme heat events. Attention-based CD methods have gained much traction because of their ability to concentrate on change regions. However, most current attention-based methods only calculate the attention matrix based on features extracted from a single image but fail to consider the correlation of features extracted from images at different times. The effectiveness of the correlation of learned features from change regions in bi-temporal images is an essential element influencing the improvement of CD performance. This paper proposes a Cascaded Attention-Induced Difference Representation Learning (CADRL) method for multispectral CD to explore the correlation of features extracted from bi-temporal images to obtain more discriminative features. The proposed CADRL method contains three modules: the feature extraction module, the Cascaded Cross-attention based Difference Learning Module (CCADLM) and the detection module. First, the feature extraction module extracts multi-scale features from bi-temporal images. Then, CCADLM generates more discriminative features by fusing difference features and the cross-attention matrix learned from the temporal attention unit from multiple levels to explore the correlation information of change regions in bi-temporal images. Finally, the learned discriminative features are fed to the detection module to gain the final detection map. Experimental results on three multispectral datasets demonstrate that the CADRL method outperforms other existing CD algorithms.