Effective monitoring of grassland biodiversity is essential, yet remains challenging because imaging spectroscopy estimates of plant diversity are highly sensitive to methodological, including scale-related, choices. Using unmanned aerial vehicle (UAV)-based imaging spectroscopy over a long-term nitrogen addition experiment in Inner Mongolia, China, we evaluated how spectral partitioning, per-band normalization, plot size, spatial resolution, and normalized difference vegetation index (NDVI) affect spectral diversity (SD) estimation of species richness. We compared four SD metrics across five plot sizes (1, 4, 9, 16, and 25 m²) and three spatial resolutions (2, 10, and 20 cm) under full-spectrum and partitioned frameworks with no, local, or global normalization, and quantified the contributions of spectral regions and NDVI. Spectral partitioned models generally outperformed full-spectrum models, but the optimal normalization was metric-specific: local normalization favored coefficient of variation (CV) and spectral variance (SV), whereas global normalization favored convex hull area (CHA) and spectral species richness (SSR). Except for SSR, local normalization also produced more stable SD–species richness relationships across plot sizes. Sensitivity to spatial coarsening of spectral data was strongest in smaller plots and weakened in larger plots, whereas CHA remained robust across different spatial resolutions. NDVI mainly improved predictions when SD alone performed poorly, especially for SV and SSR. Spectral contributions also diverged among metrics: near-infrared contributions increased with plot size for CV and SV, whereas CHA and SSR were driven mainly by visible and red-edge information. These results highlight the need to match SD metrics and processing workflows to target sampling scales for scalable grassland biodiversity monitoring.