The Spectral Variation Hypothesis (SVH) posits spectral diversity (SDiv) as a proxy for species diversity, yet its application in grasslands remains inconsistent due to confounding factors including intra- versus interspecific variation, species dominance, structural diversity, and data processing choices. Here we integrated 37,672 leaf-level hyperspectral measurements from 112 species with field inventories across three temperate steppe types along a climatic gradient in Inner Mongolia. Ten diagnostic bands (432, 586, 654, 705, 717, 754, 1452, 1712, 1928, 2304 nm) were identified via Monte Carlo simulation, and predictive performance for taxonomic and phylogenetic diversity was evaluated across two normalization methods (brightness and per-band) and three SDiv metrics: coefficient of variation (CV), convex hull area (CHA), and convex hull volume (CHV).

Selected bands matched or exceeded full-spectrum performance (e.g., brightness-normalized CHV improved the correlation with species richness (SR) from r = 0.24 to 0.42). Decomposing SDiv revealed opposing effects: interspecific variation consistently drove positive biodiversity relationships, with the strongest correlations reaching r = 0.77 for SR and r = 0.71 for Faith's phylogenetic diversity, whereas intraspecific variation acted primarily as a confounding factor for taxonomic diversity, driven especially by dominant species. Interspecific differences among dominant species accounted for the majority of the predictive power, while rare species contributed minimally to taxonomic diversity predictions, though their spectral variation showed positive associations with phylogenetic diversity in certain habitats. Normalization effects were pronounced and habitat-dependent: per-band normalization favored amplitude-sensitive metrics like CV, particularly in typical steppe, whereas brightness normalization benefited shape-sensitive metrics such as CHA and CHV, especially in meadow steppe. Typical steppe produced the strongest correlations overall (per-band CV with SR: r = 0.73), meadow steppe favored brightness-normalized metrics (interspecific CHV with SR: r = 0.72), and shrub-encroached steppe required decomposition to reveal phylogenetic signals. Structural diversity, particularly height variation, selectively modulated these relationships. Mixed-effects models confirmed community type as the dominant modulator, accounting for 54–93% of explained variance.

By hierarchically decomposing SDiv and integrating dominance, structure, and habitat context, our framework provides a mechanistic basis for canopy-scale SDiv and advances the SVH into a scalable tool for grassland biodiversity assessment.