CDAFormer: Hybrid Transformer-based contrastive domain adaptation framework for unsupervised hyperspectral change detection

Abstract

Hyperspectral image (HSI) change detection is a technique used to identify changes between HSIs captured in the same scene at different times. Most of the existing deep learning-based methods can achieve wonderful results, but it is difficult to generalize well to other HSIs with different data distributions. Moreover, it is expensive and laborious to obtain the annotated dataset for model training. To address above issues, we propose a hybrid Transformer-based contrastive domain adaptation (CDAFormer) framework for unsupervised hyperspectral change detection, which can effectively use prior information to improve the detection performance in the absence of labeled training samples by separately aligning the changed and unchanged deference features of two domains. Concretely, the difference features of the two domains are fed into the hybrid Transformer block for preliminary coarse contrastive domain alignment. Then, the positive and negative feature pairs generated from the hybrid Transformer block are prepared for the loss function level fine alignment. Particularly, the domain discrepancy can be bridged by pulling the category-consistent difference feature representation closer and pushing the category-inconsistent difference feature representation far away, so as to maintain the separability between domain invariant difference features. The acquired domain invariant distinguishing features are subsequently fed into the fully connected layers to derive the detection results. Extensive experiments on widely used datasets show that the proposed method can achieve superior performance compared with other state-of-the-art methods.