A Dual Attention Fusion Network for SAR-Optical Land Use Classification Based on Semantic Balance

The rapid development of remote sensing technology has enhanced the acquisition of multi-modal data. Optical and synthetic aperture radar (SAR) images, as a uniquely complementary pair, have been proven highly effective in land use classification tasks, allowing for more accurate and detailed analysis of surface characteristics. However, existing approaches often rely on simple fusion methods and ignore the semantic relationships between modalities. Moreover, network models often focus on merging extensive contextual semantic information across channels to capture broader scene features, neglecting the connection between spatial textures and details. To address these issues, the dual attention optical and SAR image fusion network (DAOSNet) is proposed in this paper to balance semantic interactions between modalities and supplement local spatial information. DAOSN et consists of two key modules. Firstly, the dual attention fusion Transformer (DAFT) module can create cross-modal sensory fields to enhance global interactions between pixels in the spatial dimension. Additionally, the modal semantic balance (MSB) loss is introduced to promote deeper semantic collaboration between modalities and preserve the unique perspectives of each modality. The experimental results on the WHU-OPT -SAR dataset demonstrate that the proposed approach outperforms other fusion-based approaches, with an improvement of 1.4% in overall accuracy (OA), 10.4% in mean intersection over union (MIoU), and 1.8% in Kappa compared to the MCANet.