Prediction of upland rice yield under drought stress using machine learning

Upland rice farming in Nagaland faces erratic rainfall and prolonged drought that severely limit productivity. We hypothesized that readily measurable phenotypic traits can reliably predict grain yield without molecular data. This study aimed to compare predictive performance of machine learning models; identify key phenotypic traits influencing yield; and develop a practical selection framework for regional breeding programs operating under resource constraints. Twenty eight indigenous upland rice (Oryza sativa L.) genotypes were evaluated under rainfed, drought conditions during the 2023–2024 kharif season using a completely randomized block design with two replications. Eighteen phenotypic traits encompassing root architecture, shoot morphology, phenological development and yield components were measured. Three supervised machine learning approaches-multiple linear regression (MLR), LASSO regression and Random Forest (RF)-were applied to predict grain yield per plant (GYPP). Specific root length (SRL), root dry weight (RDW) and test weight (TW) emerged as significant yield predictors across all three models. The MLR model achieved the highest accuracy (R² = 0.986; root mean square error [RMSE] = 0.296), while LASSO demonstrated superior generalization (cross-validated R² = 0.874; CV-RMSE = 0.342). Random Forest confirmed findings through nonlinear analysis (R² = 0.729). The convergence of traits across models validates their robustness. Phenotypic trait-based prediction provides a cost-effective approach for selecting drought-resilient genotypes without molecular data, offering practical value for resource-limited breeding programs in Nagaland and similar rainfed environments.