Remote sensing images change detection based on MCRASN

In order to improve the accuracy of change detection in co-registered high-resolution remote sensing images, a Siamese network combining mobile convolution and relative attention (MCRASN) was proposed based on ChangeFormer. A multi-stage combined encoder was constructed to replace the original network encoder by using vertical layout combined with mobile convolution and relative attention to efficiently capture the required multi-scale detailed features and pixel correlation information, and the difference module was improved to be a learnable distance metric module for distance calculation. At the same time, the equalized focal loss (EFL) loss function was introduced to solve the problem of imbalance between positive and negative samples in the dataset to achieve accurate change detection. The experimental results show that the proposed MCRASN method has better change detection performance on the LEVIR-CD dataset, with precision, recall, F1 score and overall accuracy of 93.94%, 89.26%, 91.54% and 99.18%, respectively, which is superior to previous methods.