Prediksi Hasil Tanaman Padi Menggunakan Random Forest dengan Seleksi Fitur Berbasis Feature Importance dan Optimasi Hyperparameter GridSearchCV

Rice production in Indonesia fluctuates due to agronomic, spatial, and climatic factors. This study develops a rice production validation model using Random Forest Regressor with feature importance-based feature selection and GridSearchCV hyperparameter optimization. Agricultural data from BPS for 2018-2024 were combined with annual NASA POWER weather data from 37 provinces in Indonesia. The model predicts rice productivity and converts the prediction into production using actual harvested area. Feature selection reduced 32 initial predictors to 16 final features. The optimized model achieved production R² of 99.79%, adjusted R² of 99.73%, RMSE of 115,600.28 tons, MAE of 61,194.75 tons, MAPE of 6.95%, and SMAPE of 6.86%. It is important to note that the high production R² of 99.79% is substantially driven by the mathematical dominance of actual harvested area as a multiplier in the production conversion formula, rather than solely reflecting the predictive power of the climate model. The core predictive performance of the model at the productivity level (R² = 75.45%, MAPE = 6.95%) more accurately represents the model’s generalization capability. Research limitations include limited historical data for newly established provinces in Papua and the relatively short study period of 2018-2024. These results indicate that the proposed method provides accurate validation for rice production analysis.