Dictionary Attention-Weighted Cross-Domain Contrastive Learning for Remote Sensing Image Change Detection

Abstract

Significant progress has been made in remote sensing image change detection due to the rapid development of deep learning techniques. Convolutional neural networks (CNNs) have become foundational models in this field. Previous works on remote sensing image change detection has utilized domain adaptation methods, achieving promising predictive performance. However, the transferable knowledge between source and target domain has not been fully exploited. In this paper, we propose a novel cross-domain contrastive learning approach for remote sensing image change detection, which correlates source and target domain using contrastive princi-ples. Specifically, we introduce a transferable cross-domain dictionary learning scheme where a shared dictionary between the source and target domains generates sparse representations. Based on these representations, we compute attention weights and propose an attention-weighted con-trastive loss to enhance knowledge transfer between source and target domains. Experiments demonstrate the effectiveness of the proposed methods on public remote sensing image change detection datasets.