Meng et al. (2025) Can machine-learning methods better characterize the relationships between crop yields and water disaster intensity at different growth stages?

Identification

• Journal: Field Crops Research
• Year: 2025
• Date: 2025-11-15
• Authors: Huayue Meng, Long Qian, Rangjian Qiu
• DOI: 10.1016/j.fcr.2025.110231

Research Groups

• State Key Laboratory of Water Resources Engineering and Management, School of Water Resources and Hydropower Engineering, Wuhan University, Wuhan 430072, China
• Changjiang Institute of Technology, Wuhan 430212, China

Short Summary

This study investigates whether machine learning methods (Random Forest and XGBoost) can better characterize the nonlinear relationships between major crop yields and water disaster intensities across different growth stages compared to multiple linear regression. The findings demonstrate that machine learning models significantly outperform linear models in accuracy, especially for individual growth stages and severe disaster scenarios, offering improved insights for agricultural disaster assessment.

Objective

• To examine if machine-learning methods can better characterize the relationships between crop yields and water disaster intensity at different growth stages compared to traditional linear models.

Study Configuration

• Spatial Scale: Middle-lower reaches of Yangtze River
• Temporal Scale: 1990–2020 (31 years)

Methodology and Data

• Models used: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Multiple Linear Regression (MLR) for comparison.
• Data sources: Major crop yields (from Statistical Yearbook), drought and flooding intensities (quantified by meteorological data).

Main Results

• The coefficient of determination (R²) values for RF and XGBoost significantly exceeded those of MLR for both the whole growth period (0.38–0.39 vs. 0.24) and individual growth stages (0.42–0.43 vs. 0.28).
• Root Mean Squared Error (RMSE) values of machine learning methods were smaller than MLR in most cases, particularly for individual growth stages.
• Improvements from machine learning methods were more pronounced in heavy-disaster cases.
• The three methods agreed on the most impactful water disasters in 80% of cases for the whole growth period and 55% for individual growth stages.
• Machine learning methods were more sensitive than MLR in detecting drought impacts.
• RF and XGBoost consistently identified the single most impactful water disaster in 95% of cases, but the top three were consistently identified in only 60% of cases.
• Dominating water disasters identified for specific crops were: flowering flooding for cotton; flowering flooding and flowering drought for oilseed; tillering flooding and tillering drought for wheat; and water disasters at all stages for maize.

Contributions

• Provides new insights into evaluating agricultural water disaster impacts under climate change by demonstrating the superior performance of machine learning methods.
• Highlights the effectiveness of Random Forest and XGBoost in capturing nonlinear relationships between crop yields and water disaster intensity at different growth stages, outperforming traditional multiple linear regression.
• Reveals the increased sensitivity of machine learning methods in detecting drought impacts and their improved accuracy in heavy-disaster scenarios.

Funding

• Not specified in the provided text.

Citation

@article{Meng2025Can,
  author = {Meng, Huayue and Qian, Long and Qiu, Rangjian},
  title = {Can machine-learning methods better characterize the relationships between crop yields and water disaster intensity at different growth stages?},
  journal = {Field Crops Research},
  year = {2025},
  doi = {10.1016/j.fcr.2025.110231},
  url = {https://doi.org/10.1016/j.fcr.2025.110231}
}