Testing the accuracy and transferability of remotely sensed biomass models across heterogeneous grasslands

Grassland aboveground biomass provides key insights into ecological processes such as carbon sequestration, animal movement patterns, and agricultural management practices. Different model types have been developed to estimate grassland biomass from satellite imagery. However, differences in model performance across sites with varying management and ecology remain largely understudied. In this study, we compared accuracy and transferability of empirical, physically-based, and hybrid models to estimate grassland biomass from multispectral Sentinel-2 data in an agnostic scenario, i.e., the models were not provided with any site-specific information beyond the spectral data. Based on field data from five study sites in Europe and the United States, we assessed (1)site-level accuracy of biomass estimation models, (2) model transferability between sites (domain shift), (3) the performance of models trained or optimized with data from multiple study sites (domain generalization), and (4) the relationship between epistemic uncertainty and model transferability. Our results showed that (1) all models exhibited comparable performance at the site level, (2) physically-based models showed the highest degree of transferability between sites, (3) no model consistently outperformed all other models when trained or optimized with field data from multiple sites, and (4) epistemic uncertainty was not necessarily a reliable measure of model applicability to unseen data. Our findings demonstrate the challenges associated with grassland biomass models under domain shift. This elucidates limits to agnostic inference in targeting diverse grasslands and highlights that model transferability is an integral part of performance assessment towards scalable satellite-based grassland monitoring systems, especially as the community increasingly deploys models at continental to global scales.