AWDA: Adversarial and Weighted Domain Adaptation for cross-dataset change detection

Abstract

Recent advancements in change detection (CD) using fully-supervised methods have been significant; however, effectively applying CD in scenarios where labels are unavailable remains a challenge. To address this, our study introduces a new task, domain adaptive change detection (DACD), which transfers change knowledge from a labeled CD dataset (source domain) to an unlabeled CD dataset (target domain). In practice, two challenges hinder change knowledge transfer across domains: domain shifts, such as resolution differences and change semantic discrepancies, and imbalanced distribution between the minority change class and the dominant no-change class. To tackle these issues, we propose a novel Adversarial and Weighted Domain Adaptation (AWDA) framework for DACD. AWDA employs a Siamese encoder–decoder network shared between source and target domains to extract features and make predictions from bi-temporal remote sensing images. Moreover, AWDA incorporates three cross-domain learning strategies for learning domain-invariant CD representations: (1) supervised learning, which uses all the labeled data of the source domain to train the model to obtain initial CD capability, (2) domain adversarial training, which aligns the features between the source and target domains adversarially, and (3) class-weighted self-training, which dynamically computes and assigns class weights for the self-training on the unlabeled data of the target domain. The proposed AWDA effectively mitigates cross-domain shifts and preserves the integrity of the minor change class during knowledge transfer. To evaluate our method’s effectiveness, we conducted comprehensive experiments across four cross-domain CD scenarios using three well-known building CD datasets. The results demonstrate AWDA substantially enhances CD performance in the target domain, achieving IoU increase ranging from 13.64 to 34.73, and significantly surpassing several competing domain adaptation methods. Our code will be available at https://github.com/zxt9/AWDA.