Remote Sensing Image Change Detection Based on Attention and Convolutional Neural Network

Abstract

One of the basic tasks of remote sensing image processing is change detection (CD). It aims to determine the change information in the region of interest and filter out the irrelevant change information. Existing methods are still inadequate in terms of depth feature extraction efficiency and lack robustness to pseudo change information. To overcome this problem, a multi-attention-based CD method is proposed in this paper. The proposed method optimizes and improves the existing feature extraction network, and introduces an attention mechanism in the feature extraction process, which can better extract the image feature information. A cascaded multi-attention module is designed, which can capture both remote dependencies and spatial dependencies to obtain a higher quality feature representation. The feature extraction and detection performance of the model is improved. In addition, a serious problem in CD is sample imbalance, and to overcome this problem, a batch-balance contrastive loss function is introduced, which uses batch weights to modify the class weights of the original contrast loss. Compared with other methods, the proposed method improves the F1 scores by 5.3% and 5.6% on the three publicly available benchmark datasets CDD and LEVIR-CD datasets, respectively. The experimental results show that the proposed method achieves state-of-the-art (SOTA) CD performance.