Hyperspectral image classification (HSIC) serves as a fundamental task in remote sensing, enabling fine-grained and detailed interpretation of land-cover characteristics. However, substantial spectral redundancy and spatial heterogeneity make it difficult to learn discriminative spectral--spatial representations. Most existing methods exploit first-order spectral information and isotropic spatial interactions, which are insufficient for characterizing high-order inter-channel dependencies and direction-sensitive spatial structures. To overcome these issues, we propose a Manifold-Aware Spectral Covariance and Directional Anchor Network (MSCANet), which integrates second-order Riemannian-aware spectral modeling with anisotropic global dependency learning. Specifically, we first propose the Cascaded Riemannian Covariance Enhancement (CRCE) module to construct compact symmetric positive definite (SPD) descriptors from grouped spectral features. Through low-rank modeling and geometry-consistent refinement on the SPD manifold, CRCE effectively exploits intrinsic second-order correlations to amplify spectral separability, while mitigating the cubic channel-wise complexity. Second, we introduce a Spatial-Spectral Anisotropic Gaussian Anchor Attention (S2AGA), which extracts content-adaptive anchors via spatial elliptical kernels and channel-wise anisotropy to reconstruct global dependencies through degree-calibrated Markov restoration, scaling efficiently with O(NK) complexity. Experimental results on four benchmark datasets (IP, SA, WHU-HC, and WHU-HH) reveal that MSCANet achieves state-of-the-art performance, yielding overall accuracies of 97.81%, 98.50%, 98.12%, and 98.71%, respectively, while maintaining fewer parameters and lower computational cost.