A Multiview Difference Feature Perception Network for Change Detection in Bi-Temporal Remote Sensing Images

Perceiving difference features from bi-temporal remote sensing images is one of the critical steps in deep learning-based change detection (CD) models. Current methods usually fuse features of bi-temporal images from a single view at a fixed stage. However, they face two key challenges: 1) the bias of a single view leads to inadequate capture of change information and 2) the fusion strategy in the fixed stage ignores the potential correlation between the high-level semantic features and the underlying detailed features. To address these limitations, this article proposes a multiview difference feature perception network. It captures the essential difference information from various perspectives via the multiview perception module (MVPM), which runs through the encoding and decoding stages. The MVPM consists of three stages: interaction, extraction, and fusion. Specifically, the multiview interaction convolution operator is designed to obtain a unified representation of the multiview difference features during the interaction phase. Then, in the extraction stage, the unified representation is continuously decomposed into multiview difference features through the feature decoupling module (FDM). Subsequently, the multiview fusion module (MVFM) is constructed during the fusion stage, and this module can enhance the ability of neural networks to recognize change information by aggregating the multiview difference features of different receptive fields. Furthermore, an independent fusion module is designed at the end of the network to enhance the network’s perception of various scale change targets. Extensive experiments on three publicly available CD datasets validate the effectiveness of the proposed method.