Change Detection Needs Neighborhood Interaction in Transformer

Abstract

Remote sensing image change detection (CD) is an essential technique for analyzing surface changes from co-registered images of different time periods. The main challenge in CD is to identify the alterations that the user intends to emphasize, while excluding pseudo-changes caused by external factors. Recent advancements in deep learning and image change detection have shown remarkable performance with ConvNet-based and Transformer-based techniques. However, ConvNet-based methods are limited by the local receptive fields of convolutional kernels that cannot effectively capture the change features in spatial–temporal information, while Transformer-based CD models need to be driven by a large amount of data due to the lack of inductive biases, and at the same time need to bear the costly computational complexity brought by self-attention. To address these challenges, we propose a Transformer-based Siamese network structure called BTNIFormer. It incorporates a sparse attention mechanism called Dilated Neighborhood Attention (DiNA), which localizes the attention range of each pixel to its neighboring context. Extensive experiments conducted on two publicly available datasets demonstrate the benefits of our proposed innovation. Compared to the most competitive recent Transformer-based approaches, our method achieves a significant 12.00% improvement in IoU while reducing computational costs by half. This provides a promising solution for further development of the Transformer structure in CD tasks.