A Modular Agroclimatic Feature Engineering Framework for Country-Level Crop Yield Prediction Using XGBoost and LightGBM
TL;DR
This framework provides a reliable benchmark for leakage-controlled agricultural machine learning research rather than direct farm-level operational forecasting and demonstrates that engineered agroclimatic features can significantly improve country-level crop yield prediction, but model performance depends strongly on the validation setting.
Abstract
Accurate crop yield prediction is essential for supporting food-security assessment, agricultural planning, and policy-level decision-making. This study proposes a modular agroclimatic feature engineering framework for country-level crop yield prediction using XGBoost and LightGBM. The proposed framework integrates climate indicators, climate–pesticide interactions, temporal descriptors, composite agroclimatic indices, and nonlinear transformations to improve predictive representation while controlling data leakage. Experiments were conducted on 28,151 country–crop–year observations from 98 countries covering 1990–2013. To evaluate temporal and geographic generalization, models were assessed using time-based validation, Random KFold, GroupKFold by country, bootstrap confidence intervals, and held-out-country evaluation. Results show that LightGBM with the S3 feature configuration achieved the best temporal prediction performance, obtaining an R² of 0.9492 and RMSE of 21,257 hg/ha. However, held-out-country evaluation revealed lower transferability, with the best configuration achieving R² of 0.6694, highlighting the challenge of geographic generalization. The findings demonstrate that engineered agroclimatic features can significantly improve country-level crop yield prediction, but model performance depends strongly on the validation setting. This framework provides a reliable benchmark for leakage-controlled agricultural machine learning research rather than direct farm-level operational forecasting.