Remote sensing of grassland biodiversity is constrained by the inconsistent performance of the Spectral Variation Hypothesis and an optical detectability bias against sub-canopy and rare species. Here, we developed a deep learning (DL) model, Domain Knowledge-infused Cross-Attention Transformer (DK-CAT), and applied it to 2 cm hyperspectral imagery from a long-term grazing and fencing site in a temperate grassland to map the sunlight-exposed species. From the resulting species maps, we derived species richness (SR) and dominance metrics, the proportions of the mapped canopy area occupied by the top-ranked species, at three spatial scales (0.5, 1, and 1.5 m) and related them to observed SR, with spectral clustering (SC) as the unsupervised benchmark. DK-CAT classified these species with 90% accuracy. SR inferred directly from the maps corresponded more closely to the sunlight-exposed SR (adjusted R2 up to 0.51) than to full-community SR (≤ 0.38), and SC was weak for both (≤ 0.17). The dominance metrics outperformed direct inference, the SC metrics, and three spectral diversity indices (≤ 0.18) in predicting full-community SR. Their predictive performance increased with scale, from adjusted R2 of 0.21 to 0.67 under fencing and 0.34 to 0.49 under grazing, and persisted beyond the plot where direct inference declined. At the plot scale, DL dominance metrics retained most of their predictive performance in spatial mixed models and leave-one-unit-out cross-validation across management units (R2 = 0.47), whereas neither SC nor the spectral diversity indices did. The two approaches depended on distinct ecological factors, DL on high canopy height variation and few rare species and SC on low spectral variation within species. These findings identify the dominance structure of sunlight-exposed species as a robust and ecologically interpretable proxy for grassland plant diversity with potential for upscaling to spaceborne imaging spectroscopy.