Predicting Lentil Seed Yield from Agronomic Traits Using a Foundation-Augmented Machine Learning Ensemble

Abstract

Accurate agricultural yield prediction is crucial for crop management and data-driven decision-making. This study proposes a machine learning framework for predicting total lentil seed yield using 264 observations collected during a single growing season at one location in Sivas, Türkiye. Seven regression models and the tabular foundation model TabPFN were evaluated through nested cross-validation. The fixed post hoc Foundation-Augmented Averaging Ensemble (FAAE) combined TabPFN, Extra Trees, CatBoost, and XGBoost using equal weights, while the primary fully nested evaluation selected the four ensemble members within each outer training fold. The fully nested FAAE achieved R2 = 0.891 ± 0.045, comparable to the strongest individual models, while the fixed post hoc FAAE yielded R2 = 0.897 ± 0.038. Split conformal prediction produced well-calibrated uncertainty intervals at 80%, 90%, and 95% coverage. SHAP and LOCO identified pods per plant and biological yield as the dominant predictors of lentil yield. Overall, these findings indicate that FAAE can provide competitive predictive performance within a leakage-controlled framework while also supporting uncertainty quantification and biologically meaningful interpretation. Because the dominant predictors, pods per plant and biological yield, are measured at or near physiological maturity, the proposed framework is more suitable for end-of-season yield estimation than for early-season forecasting. The framework therefore shows potential for within-trial yield assessment, although validation using replicated material across additional years and locations is required before broader generalization.