Comparative Performance of SARIMA, Decision Tree, and Hybrid Models for Climate-Informed Rice Production Forecasting

Abstract

Rice production is essential to national food security, yet climate-related fluctuations create uncertainty in agricultural planning. This study compared Seasonal Autoregressive Integrated Moving Average (SARIMA), Decision Tree, and hybrid SARIMA--Decision Tree models for forecasting monthly rice production in Bojonegoro, Lamongan, and Ngawi Regencies. The dataset comprised 84 monthly observations for each regency from January 2018 to December 2024, including rice production, rainfall, temperature, and sunshine duration. The procedure included Augmented Dickey--Fuller testing, seasonal decomposition, SARIMA parameter optimization, Decision Tree tuning, and residual correction, in which climate variables and temporal features were used to model SARIMA errors. The dominant predictors were sunshine duration in Bojonegoro and temperature in Lamongan and Ngawi. The hybrid model produced the lowest mean absolute percentage error in all three regencies: 32.23% in Ngawi, 36.76% in Bojonegoro, and 46.18% in Lamongan. Although the hybrid approach consistently improved upon the standalone models, the remaining errors indicate limited absolute forecasting accuracy, particularly in Lamongan. The findings support the residual-based integration of seasonal and non-linear models while highlighting the need for richer, location-specific climate and agricultural data.