Abstract:
Rice (Oryza sativa) is a primary food source for over half the global population, yet its yield is increasingly threatened by climate-induced microclimatic variations, particularly during sensitive growth stages like flowering. This study presents the development of a Machine Learning (ML)-based Early Warning System (EWS) designed to predict adverse canopy microclimate conditions, specifically temperature, humidity, and wind that influence rice yield and spikelet sterility. Utilizing high-resolution microclimatic data from the BG-358 rice canopy monitoring system, a locally developed IoT-based platform for real-time monitoring of microclimate, data were collected at 2-minute intervals from 2019 to 2024, resulting in a dataset of approximately 200,000 data points for two flowering seasons (each nearly 1 month) per year at RRDI Bathalagoda. These data, combined with NASA global forecast inputs, were split into 70% for training and 30% for testing, ensuring robust evaluation and reliable predictions to support proactive agricultural decision-making. The study employed Random Forest, Support Vector Machines, and Neural Networks to predict canopy temperature, relative humidity (RH), and wind conditions. After data pre-processing, including normalization and temporal feature engineering, models were evaluated using MAE, RMSE, and R² metrics. The Random Forest model demonstrated superior performance with an MAE of 0.78 °C, RMSE of 1.15 °C, and R² of 0.91 for temperature prediction. RH and wind predictions also showed high accuracy, with RH remaining above the 60% critical threshold in most cases, though wind forecasts, especially above the 20 km/h threshold, exhibited greater variability due to localized microclimatic conditions. Integrating real-time canopy data with weather forecasts enhanced prediction robustness, enabling heat-induced spikelet sterility anticipation up to 48 hours in advance with 85% reliability. The developed EWS incorporates real-time dashboards and automated alerts, demonstrating strong potential in mitigating climate-related rice yield losses and supporting sustainable agriculture. Future work will expand datasets across varieties and regions and foster collaboration with agricultural stakeholders for wider adoption and broader-scale yield forecasting to strengthen food security planning.
