Scene Classification of Remote Sensing Images Based on Improved HrNet

Since remote sensing images are acquired by long-distance aerial photography, they contain complex and variable feature targets, and there will be occlusion, proximity, and scale inequality between different targets, so it is very important to choose suitable feature information to express the image when performing scene classification of remote sensing images. Therefore, we improved HrNet and proposed the ASFI-HrNet model. We added Acmix to enable the model to adaptively capture information from different regions; added IEBN to enable the model to reduce the effect of noise; added four horizontal convolutional branches to integrate information from different scales to a greater extent; and added SA-based feature fusion module to further enhance feature extraction. Through a series of comparison experiments and ablation experiments, it is found that the accuracy of the ASFI-HrNet model on the RSSCN dataset is 96.43%, which is 3.4% higher than that of the HrNet model, while ASFI-HrNet obtains higher accuracy and various evaluation metrics of image classification than the other classification models in the experiments.