Abstract
Unlike natural images, remote sensing images exhibit significant spatial complexity and minimal intra-class differences, presenting considerable challenges in the field of remote sensing scene image classification (RSSC). Although existing convolutional neural networks have achieved some progress in this domain, they often fail to fully account for the unique characteristics of remote sensing images. Additionally, these networks typically suffer from excessive parameter redundancy, resulting in substantial computational burdens. This is particularly problematic given the difficulty in obtaining and labeling remote sensing data. To address these issues, this paper proposes a lightweight method (AEBANet) featuring an attention branching structure specifically designed for RSSC. First, we construct an overall feature extraction framework based on depth-wise separable convolution (DS-Conv) to ensure efficient feature extraction while maintaining accuracy. Then, we propose the Adaptive Enhanced Branch Attention (AEBA) module, a lightweight structural design that enhances the model's capability to capture key features in both channel and spatial domains. Second, we develop the Multi-Level Feature Fusion (MLFF) module to integrate features at different levels, thereby improving information flow between features and utilizing detailed shallow information to supervise the deep global information. Finally, the proposed AEBANet achieves the highest overall accuracy of 93.12%, 96.76%, and 99.52% on the NWPU, AID, and UCM datasets, respectively. Ablation studies on these datasets validate the effectiveness and necessity of each module. Additionally, the proposed method is characterized by low complexity and computational cost.