Optimized Weighted Fusion for Rice Productivity Prediction Using Multi-Source Agricultural Data

Abstract

Accurate rice productivity prediction is essential for agricultural planning, food-security monitoring and climate-risk mitigation. In this manuscript, rice productivity refers to the average harvested paddy yield per unit harvested area at the district/city-year level, expressed as quintal per hectare (q ha-1). This study proposes a validation-based optimized weighted fusion model that integrates Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Long Short-Term Memory (LSTM) for rice productivity prediction using multi-source agricultural data from districts/cities in Central Java, Indonesia, during 2010-2025. The novelty of this approach lies in its constrained, non-negative weight optimization using a separate validation period, its controlled comparison against standalone, late-fusion, stacking and boosting-based alternatives and its robustness assessment under explicitly defined climate categories. Predictors included agricultural area, climate, soil-land, agronomic and engineered temporal rainfall features, while rice production was excluded to avoid target leakage. Data from 2010-2020 were used for training, 2021-2022 for validation and fusion-weight selection and 2023-2025 for independent future-period testing. The selected weights were 0.30 for RF, 0.60 for XGBoost and 0.10 for LSTM. The optimized fusion achieved the lowest RMSE (4.2847) and the highest R² (0.2812), while XGBoost remained slightly better in MAE and MAPE. Robustness analysis showed lower error under normal climate conditions and acceptable relative error under combined extreme conditions. SHAP analysis identified drainage type, soil type, year, sunshine duration, harvested area and rainfall lag features as key predictors. The framework improves squared-error stability and provides interpretable evidence for agricultural decision support.