Machine learning to incorporate root morphology with UAV multispectral imaging for yield and nitrogen prediction in cereals

Abstract

Early-season prediction of yield and nitrogen‐related performance is essential for enabling timely agronomic interventions yet remains challenging in crops with limited prior digital phenotyping research, such as Tritordeum . Root traits, although fundamental to early nutrient uptake and crop establishment, remain largely absent in ML prediction frameworks. This study evaluated how root morphological traits, in combination with UAV-derived multispectral indices and proximal agronomic measurements, contribute to predicting yield and nitrogen efficiency indices under Mediterranean field conditions. Measurements were collected during the first three phenological stages, and a diverse set of machine-learning algorithms, including Random Forest, ExtraTrees, AdaBoost, Support Vector Regression, regularised linear models and the foundation model TabPFN, were benchmarked using a rigorous 10×5 cross-validation scheme. Yield emerged as the most reliably predictable trait, reaching an R² of 0.90 in the best multivariate configuration, while nitrogen-efficiency indices (NUE, NHI, NUtE) showed substantially higher variability and limited early-season predictability. Root diameter at the tillering stage consistently ranked among the most informative predictors, and its combination with SPAD at stem elongation, MCARI at tillering, or plant height at tillering produced the strongest yield models. These findings highlight the importance of integrating early-season below-ground information with spectral and agronomic traits to enhance prediction accuracy. Overall, the study demonstrates that accurate early-season yield forecasting in Tritordeum can be achieved using a minimal set of measurements, supporting cost-efficient monitoring and enabling actionable in-season adjustments to nitrogen management. The results also show the potential of foundation models such as TabPFN for limited agronomic datasets, providing a basis for developing scalable, data-driven decision-support tools for sustainable cereal production.

Plant phenotyping relevance

UAVマルチスペクトル、根形態・農学測定を統合した機械学習による収量・窒素関連形質の推定を中心に、複数モデルを厳密にベンチマークしており、形質推定ワークフローが実質的な方法貢献である。

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.