Transformer With Feature Interaction and Fusion for Remote Sensing Image Change Detection

With the rapid development of deep learning (DL), change detection (CD) in remote sensing (RS) image has achieved remarkable success. Nevertheless, as the image resolution improves, the visual features extracted by current methods have limited expression ability, and the networks generally suffer from spatial degradation, both of which lead to incomplete boundary detection and the undetection problem of small changed areas. At the same time, the registration errors in image pairs make remote sensing image change detection (RSCD) more challenging. To alleviate the aforementioned issues, this article proposes a transformer with feature interaction and fusion network (TFIFNet) for CD. To be specific, the proposed network utilizes the advantages of transformer in long-range dependence modeling first, which can learn feature representations with spatial-temporal information from a global perspective. Then, to alleviate the irrelevant changes caused by image registration errors, the bitemporal feature interaction module (BFIM) is proposed, which utilizes an attention mechanism to learn the bitemporal background distribution. Subsequently, an intertemporal joint-attention (JointAtt) is introduced to learn the consistency of bitemporal features for further refinement. Finally, to address the issue caused by spatial degradation during the process of network training, a triple feature fusion module (TFFM) is proposed. This module can learn spatial information from adjacent layer's features as additional spatial information. Extensive experimental studies show that the proposed network achieves the most advanced results on two CD benchmark datasets.