Solar Radiation Prediction using XGBoost across Topographic Zones: ENSO Variability and Implications for Chili Productivity in Yogyakarta 10.32526/ennrj/25/20260110

Abstract

Chili pepper productivity in Indonesia is highly sensitive to climate variability, particularly solar radiation. This study aims to predict solar radiation to optimize chili pepper productivity across four distinct topographic zones in the Special Region of Yogyakarta. We employed the Extreme Gradient Boosting (XGBoost) algorithm to predict solar radiation 18 months ahead. Models were trained using monthly ERA5 data, with the solar radiation parameter specifically validated against ground-based measurements, resulting in an average Pearson correlation (r) of 0.79. Furthermore, we analyzed the impact of the El Niño-Southern Oscillation (ENSO) on local solar radiation anomalies, and utilized time-lag correlation analysis to evaluate the delayed effects of prior solar radiation on chili productivity. The XGBoost model demonstrated moderate predictive accuracy across all zones, yielding a Mean Absolute Error (MAE) of 10.87-12.38 W/m², Root Mean Square Error (RMSE) of 13.61-14.92 W/m², Coefficient of Determination (R²) of 0.45-0.59, and Symmetric Mean Absolute Percentage Error (SMAPE) of 5.17-5.69%. Spatiotemporal analysis revealed that El Niño phases increased radiation intensity, whereas La Niña decreased it (average r=0.34 across topographic variations). A two-month lag effect was identified, with solar radiation from the preceding two months showing the strongest association with chili productivity among the evaluated lag periods (average r=0.32).  These findings suggest that solar radiation prediction may provide useful information for agricultural planning and climate adaptation strategies in regions influenced by climate variability.