Remote sensing image change detection based on multilevel cross-temporal fusion network

In recent years, with the rapid development of deep learning, attention modules based on fused scale features have been widely successful in remote sensing image change detection. However, the popular Softmax attention mechanism has high computational overhead and is susceptible to changes in viewing angle during imaging. To address these issues, we propose a multilevel cross-temporal fusion network that effectively balances the training efficiency and detection capability of the model while considering the reduction in computational complexity. Specifically, we propose a JointAgent attention module that unifies Softmax attention and linear attention in the deeper feature analysis of bitemporal cross-temporal images, thus achieving efficient information interaction in bitemporal images and facilitating information coupling between representations at the same level. Furthermore, we created a three-branch multilevel feature fusion structure based on difference extraction to accomplish deeper feature information interaction. The pixelwise sum branch was primarily used to improve edge information, whereas the pixelwise subtraction and channelwise concatenation branches were primarily used to acquire various features. Our models were trained on three public datasets with significantly improved efficiency and outperformed other existing models in terms of performance (such as the F1-score (F1) was improved by 1.52% and 1.93% on the Deeply Supervised Image Fusion Network dataset and Cropland Change Detection dataset, respectively).