Ablation-Based Climate-Aware Feature Engineering for Rice Productivity Prediction Using Multi-Source Agricultural Data

Climate-aware feature engineering is important for rice productivity prediction because rainfall effects may be delayed, accumulated, standardized, or expressed as extreme anomalies. This study evaluates the incremental contribution of rainfall-derived features for rice productivity prediction using multi-source agricultural data from districts/cities in Central Java, Indonesia, during 2010-2025. The analysis used 560 district/city-year observations and compared Random Forest and XGBoost under six sequential ablation scenarios: baseline features without engineered climate variables (S0), rainfall lags (S1), rainfall lags plus three-period moving average (S2), addition of a standardized rainfall anomaly index (S3), addition of rainfall anomaly (S4), and full climate-aware features including extreme dry and wet flags (S5). Rice productivity was defined as average harvested paddy yield per hectare at the district/city-year level, expressed in quintals per hectare, while rice production was excluded from the predictor set to prevent direct target leakage. Rainfall-anomaly parameters were estimated from the training period only, and categorical predictors were treated as nominal variables. The results show that climate-aware features improved XGBoost more strongly than Random Forest. For XGBoost, RMSE decreased from 4.8019 in S0 to 4.3068 in S5, representing a 10.31% reduction. The largest observed improvement occurred after adding rainfall lag features in S1, where RMSE decreased to 4.2121, and R2 increased to 0.3053. However, the differences among several top scenarios were small; therefore, scenario rankings are interpreted descriptively rather than as statistically significant superiority. The study demonstrates that ablation-based feature evaluation is necessary before adopting full climate-aware feature sets in agricultural prediction models.

@article{Kusumodestoni2026,
  title     = {Ablation-Based Climate-Aware Feature Engineering for Rice Productivity Prediction Using Multi-Source Agricultural Data},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {R. Hadapiningradja Kusumodestoni and Fatchul Arifin and Mutiara Nugraheni and Desti Setiyowati},
  doi       = {10.14569/IJACSA.2026.0170931},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170931}
}