Remote sensing image scene classification based on ResNet-18 and attention mechanism

With the advancement of remote sensing technology, the application of remote sensing image scene classification has become increasingly prevalent, and its significance continues to grow. To address the limitations of existing algorithms—such as high computational complexity, structural redundancy, and the requirement for extensive training rounds to achieve high accuracy—this paper proposes a lightweight architecture based on ResNet-18 integrated with an attention mechanism. The proposed framework utilizes a modified ResNet-18 as its backbone and incorporates efficient channel attention (ECA) and self-attention modules. By leveraging the attention mechanism, the model enhances focus on critical image information while suppressing irrelevant features, thereby improving classification accuracy. Experimental results demonstrate that within a limited number of training epochs (15 rounds), the overall classification accuracies on the public datasets RSICD and NWPU-RESISC45 reach 87.07% and 87.34%, superior to the existing algorithms such as ResNet-50, AlexNet, EfficientNetV2-S, and Swin-T. Compared with them, the accuracy rate has increased by at least 1.65% and 1.27%. Furthermore, the proposed model exhibits significantly lower complexity, with only 11.7 million parameters—substantially fewer than the above four, and the inference speed is also improved.