Pasture Biomass Monitoring in Queensland Rangelands with UAV and Satellite Cascades

Abstract

Operational satellite monitoring of pasture biomass demands models that extrapolate beyond the range of properties on which they were calibrated. We show that the dominant failure mode of Sentinel-2 pasture-biomass models in tropical Australian rangelands is the saturation and phenological inversion of greenness-based vegetation indices across sites, and that this failure can be substantially repaired with open climate and topsoil covariates from public archives. The work builds on a hierarchical pipeline that scales clip-and-weigh ground truth (n=1120 tare-corrected 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, and a seven-class deep-learning pasture mask (overall accuracy 98.6 %) to suppress mixed-pixel noise. Under a leave-one-site-out (LOSO) cross-validation protocol on twenty site-date aggregates across nine sites, spectral-only models failed to transfer across sites (R2=−0.21, RMSE=4.79 t ha-1). Appending open climate (Open-Meteo ERA5) and soil (SoilGrids 2.0) covariates, and switching to a gradient-boosted regressor on log-transformed biomass, lifted LOSO R2 to +0.07 and reduced RMSE to 4.19 t ha-1. A leaf-nitrogen growth trajectory, predicted by the TabPFN nitrogen regressor developed in our earlier pasture chemistry work, reduced LOSO error by a further 11 % relative to greenness-only growth features. Three additional 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 suggest that the residual LOSO ceiling at the current nine-property footprint is a sample-size and sensor-saturation limit rather than a feature-engineering one, and that the most tractable operational path forward is to stratify the production model by climatic zone and Queensland Land Type rather than pursue further covariates within a single global learner. Expanding UAV calibration footprints and integrating open climate, soil and plant-chemistry data are complementary, not competing, investments for operational rangeland remote sensing.

Plant phenotyping relevance

UAV・衛星画像、マスク処理、回帰モデルを組み合わせて牧草バイオマスを推定する測定パイプラインを構築し、サイト外交差検証で性能評価しているため、植物形質取得手法が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.