Predicting Indonesia's rice production using XGBoost, GAN-XGBoost, and BiLSTM-BiGRU with ENSO-IOD integration

Abstract

Food security programs require accurate rice production predictions to meet demand and plan reserves. However, incomplete historical data, limited predictor integration, and insufficient representation of large-scale climate variability hamper these predictions. Missing data require appropriate imputation, while feature selection retains relevant predictors. Indonesia spans 6° N to 11°08' S and 95° to 141° E, with diverse provincial weather conditions requiring integration with global El Niño-Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) signals. Previous studies have not combined regional weather variables, ENSO-IOD signals, data imputation, and feature selection into a rice production prediction framework. To address this gap, this research developed prediction models using extreme gradient boosting (XGBoost), generative adversarial network (GAN)-augmented XGBoost, and a hybrid bidirectional long short-term memory-bidirectional gated recurrent unit (BiLSTM-BiGRU) with explicit ENSO-IOD integration. Several imputation and feature selection strategies were evaluated. XGBoost with K-Nearest Neighbors imputation and without feature selection performed best, achieving a root mean square error (RMSE) of 0.0082, a mean absolute percentage error (MAPE) of 5.86%, and an R² of 0.9991. Neither GAN-augmented XGBoost nor BiLSTM-BiGRU outperformed XGBoost. Ablation tests showed that ENSO-IOD integration reduced RMSE from 0.0182 to 0.0082. These findings can support rice availability estimation, early production shortfall identification, and rice reserve requirement decisions.