Basmati Rice Yield Prediction Using Machine Learning Techniques

Accurate pre-harvest estimation of rice yield at district and state levels is essential for national production forecasting, evidence-based policymaking, and enhancing farmers’ incomes through market-aligned strategies. Traditional methods, such as field surveys and farmer interviews, are widely used but are often limited by subjectivity, high costs, and delayed outputs. Recent advancements in RS and GIS–based geospatial technologies, combined with crop growth models, offer scalable, cost-effective, and timely alternatives for yield forecasting. This chapter reviews different rice/aromatic Basmati rice yield models being used and also presents a case study evaluating the use of empirical and machine learning models, moderate multiple linear regression (MMLR), random forest (RF), and extreme gradient boosting (XGBoost), for pre-harvest yield estimation of Basmati rice in Karnal district of Haryana. The modelling was based exclusively on weather variables, supported by the agroecological uniformity of the Indo-Gangetic Plains and the prevalence of assured irrigation across the Basmati-growing area. This ensures that weather parameters serve as the primary drivers of yield variability. MMLR delivered reasonable accuracy (RMSE and MAPE within 15%) but was limited by its linear structure. RF improved performance by capturing nonlinear interactions, while XGBoost consistently outperformed both, across long-duration, traditional, and short-duration varietal groups, using 17 years (2007–2023) of historical data. Notably, XGBoost demonstrated strong generalization, accurately forecasting yields for the unseen year 2024. It achieved optimal bias–variance trade-offs for long-duration Basmati and perfect Test https://www.w3.org/1998/Math/MathML" display="inline"> R 2 values for traditional types. For recently released short-duration Basmati varieties, with only 12 years of data (2012–2023), XGBoost maintained high reliability. These results establish XGBoost as a robust, generalisable, and climate-resilient tool for Basmati rice yield forecasting. Looking ahead, deep learning models, such as long short-term memory (LSTM) networks, convolutional neural networks (CNNs), and transformer-based architectures, may further enhance predictive accuracy by capturing complex, nonlinear, and temporal dynamics in yield-predictor relationships.