DPCA: Dynamic multi-prototype cross-attention for change detection unsupervised domain adaptation of remote sensing images

Abstract

Unsupervised domain adaptation (UDA) is a key technique for enhancing the generalization and reusability of remote sensing image change detection (CD) models. However, the effectiveness of UDA is often hindered by discrepancies in feature distributions and sample imbalances across disparate CD datasets. To address these issues, we propose the Dynamic Multi-Prototype Cross-Attention model for UDA in CD. This approach enhances the representation of complex land cover features by incorporating multi-prototype features into a cross-attention mechanism, while addressing sample imbalance through a novel pseudo-sample generation strategy. The Multi-prototypes and Difference Feature Cross-Attention Module iteratively updates the multi-prototype features and integrates them with a classical two-stream CD model. This allows the model to achieve domain alignment by minimizing the neighborhood distance between the global multi-prototype features and high-confidence target domain prototype features. In addition, we propose the Sample Fusion and Pasting module, that generates new target domain-style samples of changed regions to facilitate CD-UDA training. Experimental evaluations on the LEVIR, GZ, WH, and GD datasets confirm that DPCA model effectively bridges the feature distribution gap between the source and target domains, significantly improving the detection performance on the unlabeled target domain. The source code is available at https://github.com/Fanrongbo/DPCA-CD-UDA.