Abstract

The operational satellite monitoring of pasture biomass requires models that transfer beyond the properties on which they were calibrated. We present a hierarchical, open-source cascade that scales in situ clip-and-weigh biomass (n = 1120 samples across eleven sites on five Queensland properties) through UAV digital-surface-model imagery to Sentinel-2 predictions, using TabPFN—a pre-trained transformer foundation model for small tabular data—as the regressor at all three nested spatial scales. Under a leave-one-site-out (LOSO) protocol on twenty site–date aggregates across nine sites, spectral-only Sentinel-2 models failed to transfer (best R2=−0.15, RMSE 4.62 t ha−1). Appending open climate (Open-Meteo ERA5) and topsoil (SoilGrids 2.0) covariates and evaluating five learners (GBM, RF, XGBoost, TabPFN, and GBM + TabPFN stack) on log-transformed biomass increased LOSO R2 to −0.05 and reduced RMSE to 4.43 t ha−1; a leaf-nitrogen growth trajectory predicted by the TabPFN nitrogen regressor from our earlier pasture-chemistry work reduced pixel-level LOSO RMSE by a further 4%. Three alternative covariate classes—BARRA-R2 reanalysis climate, three independent fractional-cover products, and Sentinel-1 C-band SAR backscatter—were tested and rejected, all hitting the same RMSE floor. The symmetric negative results indicate that the residual LOSO ceiling on the current nine-property footprint is a sample-size and optical-saturation limit rather than a feature-engineering one; the most tractable operational path forward is to stratify the production model by climatic zone and Queensland Land Type rather than pursuing further covariates within a single global learner.

IPC Classification

Keywords