Remote Sensing Images Change Detection Using Triple Attention Mechanism to Aggregate Global and Local Features

The change detection of high-resolution images plays an important role in practical applications. However, most existing studies use local or global attention mechanisms alone to filter and screen for changing features. In this study, we proposed a triple attention multiscale fusion network (TAMFNet) that can effectively utilize both global and local attention mechanisms, thereby improving the ability to detect the location of change areas and fully outline the change areas. First, we employed a fully convolutional network to extract features from dual temporal images at different scales. Second, three complementary attention mechanisms, namely, the spatial attention mechanism (SAM), channel attention mechanism (CAM), and the multihead self-attention (MSA) module, were integrated to extract and fuse global and local features. Finally, to address semantic and scale differences, we utilized the cross scale fusion (CSF) module, pyramid pooling module (PPM), and pyramid receptive field (PRF) module to aggregate features from adjacent scales for comprehensive feature transmission. To demonstrate the effectiveness of our method, we tested it on the LEVIR-CD and WHU-CD datasets. The results showed that our model achieved intersection over union (IOU) scores of 80.17% and 77.23% on the datasets, outperforming comparative models. Ablation experiments on the LEVIR-CD dataset confirmed the positive impact of each intermediate module in TAMFNet, with an overall 2.35% increase in IOU score.