An asymmetric dual-branch modeling method was proposed to address the modality discrepancy and heterogeneous information structures in the joint classification of hyperspectral and LiDAR images. Separate feature extractors were designed for the dominant and auxiliary modalities. In the hyperspectral branch, a serial structure combining a vision transformer and a convolutional neural network was constructed. A central-focus Mamba module was introduced to enhance perception of central regions through modeling context via spiral paths. A spatial-spectral refinement module was applied to improve feature expression quality via fine-grained optimization. In the LiDAR branch, a lightweight convolutional structure was used to extract structural and elevation information, reducing redundant modeling while maintaining scale alignment. Experiments were conducted on three benchmark remote sensing datasets. Superior performance was achieved in terms of overall accuracy, average accuracy, and Kappa coefficient, demonstrating strong robustness and generalization ability. Results show that classification performance is significantly improved by the coordinated design of modality-specific modeling and region-aware enhancement mechanisms.