| Summary: | Building change detection (BCD) is crucial for urban construction and planning. The powerful discriminative ability of deep convolutions in deep learning-based BCD methods has considerably increased the accuracy and efficiency. However, dense and continuously distributed buildings contain a wide range of multi-scale features, which render current deep learning methods incapable of discriminating and incorporating multiple features effectively. In this work, we propose a Siamese cross-attention discrimination network (SCADNet) to identify complex information in bitemporal images and improve the change detection accuracy. Specifically, we first use the Siamese cross-attention (SCA) module to learn unchanged and changed feature information, combining multi-head cross-attention to improve the global validity of high-level semantic information. Second, we adapt a multi-scale feature fusion (MFF) module to integrate embedded tokens with context-rich channel transformer outputs. Then, upsampling is performed to fuse the extracted multi-scale information content to recover the original image information to the maximum extent. For information content with a large difference in contextual semantics, we perform filtering using a differential context discrimination (DCD) module, which can help the network to avoid pseudo-change occurrences. The experimental results show that the present SCADNet is able to achieve a significant change detection performance in terms of three public BCD datasets (LEVIR-CD, SYSU-CD, and WHU-CD). For these three datasets, we obtain F1 scores of 90.32%, 81.79%, and 88.62%, as well as OA values of 97.98%, 91.23%, and 98.88%, respectively.
|
|---|