Change Detection Method of High Resolution Remote Sensing Image Based on UNet Network

To address the issues of inadequate feature extraction, complicated models, and a large number of network parameters, this paper proposes a change detection method for high-resolution remote sensing images based on UNet network. Firstly, Resnet50 network is selected in the encoder part to obtain deeper features, avoid gradient disappearance and reduce training time. The scSE attention module structure is introduced in the skip connection. This module enhances the network's ability to prioritize the important channels and mine the local information better. Then, in order to improve the segmentation accuracy and obtain the context information better, atrous spatial pyramid pool is carried out in the deepest feature layer of the encoder to enrich the representation of information and obtain the multi-scale information of the target. Finally, the combination of Focal loss function and BCE loss weight was presented with improving model learning strategy. On the LEVIR-CD and CDD datasets, it is shown that proposed method has improved precision rate, recall rate and F1, and the number of model parameters is 5.37M, which is superior to other commonly used change detection networks.