Mixture of experts image transformer for change detection in bitemporal remote sensing images

Change detection identifies alterations in the state or structure of a target by analyzing observations at different timestamps. Existing methods primarily employ deep convolutions and Transformers for feature extraction, and have achieved notable success. However, for some of these methods, the absence of low-semantic information results in insensitivity to small-scale local changes, particularly in complex scenarios. And for the other methods, through Transformer models long-range dependencies in bitemporal images pairs, it remains inefficient in processing multi-scale feature fusion. To solve these problems, we propose Mixture of Experts Image Transformer (MEIT), a novel framework for efficient multi-scale feature fusion. To be specific, we design two modules, MTransformer-1 (MT1) dual-route architecture and MTransformer-2 (MT2) single-route architecture, implementing secondary weighting of tokens in general Transformers. Then, we employ MT1 to denoise preserved multi-scale feature before concatenation during the upsampling stage. Finally, after each concatenation, Skip-Transformer Fusion (STF), incorporating MT2 is applied to balance multi-scale feature, to generate a more accurate change map. Extensive experiments on three public datasets demonstrate MEIT have state-of-the-art performance in bitemporal image change detection.