Support Vector Regression-Based Prediction of Rice Production Across Provinces in Sumatra Island

Abstract

This study develops a Support Vector Regression (SVR)–based forecasting framework to model rice production across the ten provinces of Sumatra, a region whose agricultural output is highly sensitive to climate variability and land-use dynamics. Rising uncertainty in rainfall-dependent rice ecosystems underscores the need for more accurate predictive tools to support regional food-security planning. The objective of this research is to construct and evaluate a multivariate SVR model that integrates harvested area, rainfall, humidity, and temperature, while accounting for nonlinear temporal patterns and structural differences among provinces. The methodological approach includes extensive feature engineering, log-transformed SVR estimation with time-series cross-validation, a specialized year-over-year model for small and volatile provinces, and a stabilization procedure to ensure temporal consistency in the predictions. Results show that the blended–stabilized model performs strongly on the 2021–2024 test period, achieving SMAPE of 16.10%, MAE of 124,975.77, RMSE of 194,853.89, and R² of 0.9637, and generating three-year-ahead forecasts supported by bootstrap-based uncertainty intervals. These findings indicate that the proposed framework effectively captures heterogeneous production dynamics and provides reliable predictions for 2025–2027. The study concludes that SVR offers a robust and interpretable foundation for agricultural forecasting in data-limited environments, though future work should incorporate higher-frequency data, additional agronomic indicators, and hybrid machine-learning or deep-learning models to further improve long-term performance.