High-resolution remote sensing image land cover classification based on EAR-HRNetV2

Abstract

The application of land cover classification is very extensive, it can be used in land resources management, urban planning, agricultural and other fields. The multi-scale variation of targets in high-resolution remote sensing images leads to low accuracy in land cover classification tasks. In response to this problem, an improved semantic segmentation model named EAR-HRNetV2 is proposed based on HRNetV2. Firstly, the attention mechanism is added to the Bottleneck residual and Basic residual modules to obtain more efficient features. Then, the atrous spatial pyramid pooling is added to the model. And it is modified to cope with the gridding issue caused by the dilation convolution. Finally, the refinement module is used to combine features from different stages and further extract semantic features. The experiments show that on the WHDLD public dataset, compared with five semantic segmentation models, the EAR-HRNetV2 model is effective with Kappa coefficient, F1-score, overall accuracy, and mean intersection over union reaching 77.85%, 73.00%, 87.92%, and 60.28%. Compared with HRNetV2, the mean intersection over union of EAR-HRNet increases by 2.93 percentage points, the F1-score increases by 2.67 percentage points, the Kappa coefficient increases by 2.37 percentage points, and the overall accuracy increases by 1.19 percentage points. EAR-HRNetV2 network provides a feasible solution for the land cover classification.