ML Models for Yield Estimation of Rice Cultivars Using UAV Imagery (www.mdpi.com)

Researchers demonstrated accurate, UAV‑based rice yield estimation for hybrid and conventional Japonica cultivars by combining multispectral aerial imagery, vegetation indices (NDVI, GNDVI, SAVI, OSAVI, MSR, DVI, NRI, NDRE, SIPI) and machine learning. They first built inversion models to map UAV VIs to key phenotypic variables — leaf area index (LAI), chlorophyll content (CC), plant height (PH) and above‑ground biomass (AGB) — then used those phenotypes (and/or VIs directly) to predict final grain yield. Data came from a 6‑band multispectral sensor (450, 555, 660, 720, 750, 840 nm; 12‑bit, 1.2 MP; GSD ≈8.65 cm at 120 m). The study compared multiple regressors (Random Forest, XGBoost, SVR, back‑propagation neural network, and multiple linear regression) with tuned hyperparameters (e.g., RF: n_estimators=70, max_depth=10; XGBoost: n_estimators=70, lr=0.05; SVR: RBF, C=60; BPNN: 100 hidden units, 1000 epochs). Significance for AI/ML in agtech: the paper shows that phased approaches (VI → phenotype → yield) and ensemble/tree‑based models tend to outperform simple direct VIs, and that the reproductive/heading stage is the optimal window for UAV acquisition to maximize predictive power. Practical implications include scalable, non‑destructive yield forecasting for cultivar management and breeding decisions, transferable model architectures for other crops, and concrete device/model configurations practitioners can reproduce to integrate into precision‑ag pipelines.