Research on Remote Sensing Image Scene Classification Based on ResNet50 Transfer-learning and ECA Attention Mechanism

In view of the limitations of traditional convolutional neural networks in remote sensing image scene classification such as low accuracy and low precision, this paper proposes a remote sensing image scene classification method based on ResNet50 transfer learning and ECA attention mechanism to improve the classification accuracy of remote sensing images. Firstly, The network model is based on ResNet50. Secondly, Initialize the ImageNet pre-trained image classification model, Fine-tune the output of the last fully connected layer and train the parameters, Frozen the bottom weight, and transfer learning is introduced to prevent overfitting of the model. Finally, add ECA modules to the last layer of Layer1 and Layer4 for appropriate cross-channel interaction and to avoid dimensionality reduction. Experiments were carried out on remote sensing image datasets WHURS19 and AID, Compare the classification accuracy, confusion matrix and the image regions concerned by Grad-CAM thermal maps. The results show that the maximum classification accuracy of the proposed method in the two data sets respectively reaches $\mathbf{9 8. 5 \%}$ and 88.9%, which can effectively improve the accuracy and precision of remote sensing image scene classification with large intra-class differences and high inter-class similarity, and verify the feasibility and effectiveness.