A dual attention mechanism improved Swin transformer for extracting eucalyptus from GF-1 imagery

The remote sensing image segmentation of eucalyptus faces the problems of missed judgment and misjudgment due to the lack of local information caused by the complex backgrounds. In this study, a semantic segmentation method for extracting eucalyptus is proposed based on the Swin transformer network framework with the fusion of double-layer attention, which a dual attention mechanism, termed as ECA-SimAM, is constructed by the fusion of the ECA and SimAM attention modules. The SimAM attention module is introduced in the global average pooling process of the ECA attention mechanism, which can more accurately measure the difference between each feature and its mean value, and obtain more accurate attention weight values, paying more attention to the areas that play a decisive role in segmentation, and improving the ability to capture important feature information for extracting eucalyptus. A fused GF-1 image of the study area is selected as the data source to construct the eucalyptus semantic segmentation datasets, and the experiment of extracting eucalyptus is conducted by using the Swin Transformer with ECA-SimAM. Compared with Deeplabv3+, U-net, Swin Transformer and the Swin Transformer with ECA, the results show that the accuracies of the improved method in extracting eucalyptus is significantly improved, which the Accuracy, Recall, F1-Score and IoU reach the highest values of 90.52%, 96.63%, 93.48%, and 87.76%, respectively, and the edge segmentation is more detailed, resulting in the optimal segmentation results for extracting eucalyptus.