Mixed-former: multi-fusion remote sensing change detection

Abstract

ABSTRACT Although remote sensing change detection methods based on deep learning have achieved relatively excellent results, they still face many challenges, such as many difficult samples and uneven sample distribution. The current CNN method is limited by the lack of effective global spatial-temporal information, while the transformer method is unable to accomplish the task of grounded application due to the complex network structure. Therefore, we propose mixedformer, a multi-fusion siamese network based on the enhancement of spatial-temporal information fusion. In this regard, the spatial-temporal symmetry module extracts valid potential contributing feature information based on a global self-attentive mechanism over the enhanced image information sequence. Because each embedding contains other embedding information, feature information can be generated at different locations to obtain denoised local feature information and global spatial-temporal information. At the same time, we argue that features in different layers, whether shallow or deep, contribute to the final detected change results with some weight. We also believe that both shallow and bottom features contribute some weight to the final detected change results. To achieve this, we use global cross-attention in the middle layer of decoding, linking valid information from deep and shallow layers, filtering noise from upsampling of the bottom features, while refining the edge information of the changes. Extensive experiments have shown that our model achieves excellent results on all three CD datasets with an appropriate amount of computation.