Dual-Branch CNN With Enhanced Bidirectional Multiscale Sparse Cross-Attention for Remote Sensing Scene Classification

Remote sensing scene classification (RSSC) remains challenging due to large intra-class diversity and high inter-class similarity, while a single backbone often struggles to capture both global semantics and local structures. This paper proposes a lightweight dual-branch convolutional neural network (CNN) equipped with an Enhanced Bidirectional Multiscale Sparse Cross-Attention (EBiMSC) module for multiscale feature fusion. The proposed network integrates two complementary branches: a streamlined ConvNeXt-Tiny branch that enhances global contextual modeling and long-range dependency capture, and a lightweight GhostNetV3 branch that captures fine-grained textures and local details, thereby reducing confusion among visually similar categories. To strengthen cross-branch complementarity, the EBiMSC module performs bidirectional multiscale sparse cross-attention to facilitate interaction across different scales, while a parallel lightweight enhancement path supplements informative local cues. The fused features are adaptively integrated through learnable weights to obtain discriminative and consistent representations, followed by an attention-guided lightweight multilayer perceptron (MLP) head for classification. Despite its concise design, the model is computationally efficient, with 19.5 million parameters and 4.24 GFLOPs, making it suitable for real-world applications. Extensive experiments on four benchmark RSSC datasets demonstrate that the proposed method achieves superior performance compared with recent CNN-based, Transformer-based, and hybrid models, validating its effectiveness and efficiency.