Lag Aware Ensemble Learning Model for Weather-Based Forecasting of Arecanut Fruit Rot Disease

Highlights A lag-aware ensemble learning framework is proposed for forecasting arecanut fruit rot disease using weather data. Statistically validated lagged meteorological features (1–3 weeks) significantly improve prediction accuracy. Random Forest and Stacking Regressor models achieved high predictive performance (R2 > 0.95). Humidity and rainfall were identified as dominant climatic drivers influencing disease progression. The framework provides an interpretable and scalable early warning system for climate-driven disease management. Abstract. Arecanut fruit rot disease, predominantly caused by Phytophthora species, is a major biotic constraint in humid tropical regions, resulting in severe yield losses and economic impact. Accurate early prediction of disease intensity remains challenging due to complex, nonlinear, and delayed interactions between meteorological factors and disease development. This study proposes a lag-aware ensemble learning framework that integrates statistical correlation analysis with machine learning to enhance weather-based forecasting of the intensity of arecanut fruit rot disease. Weekly meteorological variables—including rainfall, field and ambient humidity, wind speed, maximum and minimum air temperature, soil temperature (5 cm and 10 cm depths), and water temperature—were collected from CPCRI meteorological stations, field sensors, and validated farmer observations across multiple disease-prone locations. Disease intensity was quantified using weekly field-observed Percent Disease Intensity (PDI) collected during the monsoon season. Because complete disease observations were unavailable across all monitored locations, missing disease intensity values were estimated using a calibration framework that integrates meteorological variables and a CPCRI-validated, agronomic, rule-based disease-risk score, resulting in a harmonized, full-location PDI dataset used for model development. Pearson and Spearman correlation analyses identified humidity and rainfall as dominant climatic drivers, while lagged Spearman correlation (1–3 weeks) revealed strong delayed effects on disease progression. These statistically validated lagged variables were explicitly incorporated into model development. Multiple models—including Linear Regression, Decision Tree, Random Forest, Gradient Boosting, Voting Regressor, and Stacking Regressor were evaluated. Results demonstrate that lag-aware models consistently outperform non-lagged counterparts, with the Random Forest Regressor achieving the highest individual performance (R 2 = 0.9545, RMSE = 1.44) and the lag-aware Stacking Regressor attaining comparable accuracy (R 2 = 0.9531, RMSE = 1.46). Overall, lag-aware ensemble models improved predictive accuracy by approximately 10% to 15%, confirming their effectiveness as reliable early warning tools for climate-driven disease management in arecanut cultivation. Keywords: Arecanut fruit rot disease, Disease intensity prediction, Ensemble learning, Lag-aware forecasting, Machine learning, Precision agriculture, Random Forest regressor, Spearman lag correlation, Stacking Regressor.