Hyperspectral imaging (HSI) and light detection and ranging (LiDAR) data provide highly complementary information in terms of fine-grained spectral characterization and three-dimensional structural description, which is crucial for accurate land-cover classification in complex scenes. However, most existing multimodal fusion methods adopt isomorphic feature extraction and weighted aggregation strategies, which fail to explicitly model the intrinsic differences between spectral sparsity in HSI and geometric sparsity in LiDAR. Moreover, multi-scale fusion is often performed without explicit competitive selection mechanisms, leading to redundant feature accumulation and insufficient responses at discriminative scales. To address these limitations, this paper proposes a Competitive Sparse-Selection Mamba Network (CSS-Mamba) for hyperspectral–LiDAR joint classification. The proposed framework integrates modality-specific sparse modeling with competitive multi-scale feature selection in a unified architecture. Specifically, a Multi-Scale Spectral–Spatial Sparse Encoding (MS-SSSE) module and a Multi-Scale Geometric Structure Sparse Encoding (MS-GSSE) module are designed to separately model the spectral sparsity of HSI and the structural sparsity of LiDAR. Learnable soft-thresholding operators are embedded into the network to impose sparsity constraints at the architectural level, enabling modality-differentiated feature representations. Furthermore, a Competitive Sparse Selection (CSS) module is introduced, which combines visual selective state-space blocks with spatial competitive selection to establish explicit competition among multi-scale features, adaptively enhancing discriminative scales while suppressing redundant ones. Extensive experiments conducted on three public benchmark datasets, including Trento, Houston 2013, and MUUFL, demonstrate that the proposed CSS-Mamba consistently outperforms representative CNN-, Transformer-, and Mamba-based multimodal fusion methods in terms of overall accuracy, average accuracy, and Kappa coefficient. The results verify the effectiveness and robustness of the proposed modality-aware sparse modeling and competitive scale selection strategy, and highlight its potential as a general and efficient framework for multimodal remote sensing data fusion.