Heterogeneous Image Change Detection Based on Dual Image Translation and Dual Contrastive Learning

Nowadays, remote sensing (RS) change detection (CD) plays an important role in Earth observation applications. Recently, the value of cross-modal CD has gradually emerged because of the complementary features in content. However, existing domain adaption-based methods are generally limited to the optical domain and suffer from imbalanced information between modalities. In this article, a novel CD method based on dual image translation and dual contrastive learning (C3D) is proposed for heterogeneous RS images, including a translation module and a CD module. First, the translation module aims to learn a comparable representation between the different domains through a C3D structure based on feature samples, which can break the consistency constraint and better solve the imbalanced information. Then the similarity metric of patches is compared by contextual features at different scales in the CD module to achieve a more accurate classification of changed and unchanged pixels. The C3D is compared with state-of-the-art methods and validated by the basic experimental results on five datasets. In addition, further experiments on the translation module were also performed to explore the effectiveness of contrastive learning in the CD task.