Enhancing ResNet50 with Depth-wise Separable Convolution Squeeze-and-Excitation and Channel Squeeze-and-Excitation Modules for Accurate Remote Sensing Image Scene Classification

Remote sensing image scene classification is crucial for urban planning and disaster monitoring. Existing methods often struggle with extracting local texture and global semantic features and exploring inter-channel relationships, limiting classification accuracy. To address these issues, we propose Res2SE50 model, an enhanced ResNet50 model integrating Depth-wise Separable Convolution Squeeze-and-Excitation (DSE) and Channel Squeeze-and-Excitation (CSE) attention mechanisms. This approach preserves spatial and global features while enhancing feature extraction through cascading CSE with DSE. Evaluated with 20% training data on the AID and NWPU-RESISC45 datasets, Res2SE50 achieved accuracies of 91.52% and 94.09%, respectively—0.79% and 1.63% higher than the latest LHNet models. Visualization analysis confirms the model’s effectiveness in distinguishing scene features, demonstrating its strong performance and feasibility for remote sensing scene classification.