Forecasting Climate Variables and Their Impact on Food Crop Productivity in Java, Indonesia Using Seasonal GSTAR and Prophet Models

This study presents a hybrid forecasting framework that integrates the Seasonal Generalized Space Time Autoregressive (S-GSTAR) model for climate variables (temperature, rainfall, humidity) and the Prophet model to project the productivity of agricultural commodities (rice, corn, soybean, mung bean, peanut, cassava, sweet potato). Data from 2011 to 2022 were used to train the models, with projections extended to 2027 for climate variables and to 2032 for agricultural productivity. Challenges included high variability in rainfall data and limited historical records, which were mitigated through seasonal lag adjustments in S-GSTAR and trend-seasonality modeling in Prophet. The S-GSTAR model demonstrated high accuracy in forecasting temperature (MAPE $0.91 \% 1.40 \%$) and humidity (MAPE $1.13 \%-1.92 \%$), but was less accurate for rainfall (MAPE 16.11%-54.21%) due to high variability. Granger causality tests revealed linear effects of temperature and humidity on several commodities, while positive Transfer Entropy values indicated nonlinear influences of temperature and humidity on all agricultural outputs. The Prophet model showed high predictive accuracy for rice (MAPE $0.00 \%-4.65 \%$) and peanut (MAPE $1.29 \%-8.21 \%$), but performed suboptimally for mung bean (MAPE 4.89% $\mathbf{8 5. 0 9 \%}$). This approach offers valuable insights for long-term food security planning.