A multilevel optical-SAR imagery fusion network for land cover classification

With the continuous expansion of remote sensing image data sources, improvements in spatiotemproal resolution, and increasing acquisition frequency, the complementary imaging mechanisms of optical and SAR images offer substantial advantages for remote sensing classification in complex scenarios. Deep learning techniques leveraging multi-modal remote sensing fusion have become essential for enhancing land use and land cover (LULC) classification performance. However, the inherent differences between modalities hinder traditional optical-SAR fusion methods from achieving effective cross-model semantic consistency, often leading to noise and redundant information during modality fusion. To address these challenges, this paper proposes an optical-SAR dual-branch multi-level feature fusion network based on joint key learning, named the Key-Value Fusion Network (KVFNet). In the feature extraction phase, KVFNet introduces a shared semantic space to guide mutual attention mechaanisms between modalities through joint key learning, achieving semantic alignment and complementary feature enhancement. Subsequently, the Adaptive Atrous Spatial Pyramid Pooling(A-ASPP) module is employed to enhance multi-scale contextual modeling. Finally, a weighted fusion strategy is implemented to improve the model's discriminative capability at both the feature and decision levels. To validate the proposed model, experiments were conducted on the WHU-OPT-SAR dataset. The results demonstrate that KVFNet surpasses current mainstream methods across multiple metrics, improving overall accuracy (OA), mean intersection over union (mIoU), and Kappa coefficient by 1.7%, 3.2%, and 2.4%, respectively. These findings indicate that the proposed approach achieves greater accuracy and robustness for land cover classification tasks.