Agronomic variables outperform multispectral indices for individual plant yield prediction in Andean quinoa

Abstract

Accurate pre-harvest yield estimation is essential for decision-making in high-altitude agriculture. This study evaluated agronomic and multispectral UAV variables for near-harvest prediction of individual quinoa grain weight, with data collected across six phenological stages to identify when predictors achieve reliable performance, under Andean conditions. A total of 374 plants were monitored across six phenological stages at Santa Ana Experimental Station (Huancayo, Peru, 3280 m a.s.l.) during 2024. OLS, Random Forest, Support Vector Machine, and Neural Network models were trained using agronomic-only (AGRO), spectral-only (IND), and combined (COMP) predictor sets, evaluated through 5-fold cross-validation reporting mean ± standard deviation. Agronomic and combined models achieved moderate performance (R² = 0.22–0.25, RPD = 1.10–1.15), suitable for relative plant ranking in breeding programs, while spectral-only models failed across all algorithms (R² ≤ 0.044, CCC ≤ 0.080), constrained by saturation, phenological decoupling, and canopy heterogeneity. Variable importance analysis confirmed that late-season structural traits dominated predictions, while spectral indices contributed marginally despite including red-edge bands. These results challenge spectral-only approaches for individual plant phenotyping in heterogeneous canopies, demonstrating that integrating simple ground measurements with UAV spectral data is essential for reliable quinoa yield estimation.