Recently, remote sensing image scenes have been increasingly widely used in monitoring the environment, exploring earth resources, and predicting natural disasters. Numerous data requirements aid in the rapid development of remote sensing image scene classification. Although the deep learning-based method has achieved decent performance in scene classification, how to effectively classify remote sensing scenes with complex backgrounds and drastic scale changes remains a great challenge in the classification task. To address this issue, this paper proposes a fine-grained approach to detect the salient region, uses the global and local branches to combine the global and local parts, and extracts the global features and local key information from the whole image and key region, respectively. To verify the effectiveness of the proposed method, comparative experiments are conducted using ResNet18 on three public remote sensing image scene classification datasets, and the experimental results show that the accuracy of the proposed method is better than that of most advanced methods.