Field spectroscopy and machine learning successfully predict grassland forage quality and quantity across climate zones

Grasslands cover one-third of Earth’s land surface and are essential for livestock forage provision. Monitoring forage biomass and quality is key for sustainable management. Hyperspectral remote sensing and field spectroscopy is promising, but global models often fail across biomes. We compiled data from temperate, humid tropical, and dry subtropical grasslands in Europe and Africa, spanning local growing seasons and management gradients. Using machine-learning models, we assessed the performance and transferability of global and regional predictions for forage quality (metabolizable energy), and quantity (aboveground biomass), and their combined proxy (metabolizable energy yield). Random forest regression performed best for metabolizable energy (nRMSE = 0.108, R$^2$ = 0.68), aboveground biomass (nRMSE = 0.145, R2 = 0.53), and metabolizable energy yield (nRMSE = 0.153, R$^2$ = 0.58). Neural networks showed highest global-to-regional transferability (nRMSE as low as 0.083), while globally trained partial least squares models outperformed regional ones (ΔnRMSE: 0.211 to 0.037). Forage quality was pre- dicted most accurately, likely due to consistent variation in plant functional traits and strong spectral correlations. In contrast, forage quantity was harder to model due to region-specific canopy structure and pigment differences. No method achieved full spatial transferability. Our findings highlight both the potential and limitations of hyperspectral models for forage monitoring, particularly the consistent accuracy of forage quality predictions and the superior performance of random forest models. Transferability across regions was only feasible when models accommodated local variability. Expanding spectral datasets, advancing sensors, and refining models may improve predictions, supporting more sustainable grassland management worldwide.