Mitigating crop modeling uncertainties through machine learning in drylands

Area of Science:

• Agricultural Science
• Climate Science
• Data Science

Background:

• Dry farming systems in drylands are highly susceptible to climate extremes, leading to significant yield variability and threatening food security.
• Reliable crop modeling is crucial for managing these systems, especially in data-scarce regions.

Purpose of the Study:

• To evaluate and improve the performance of the CSM-CERES-Wheat crop model in Iranian drylands using various meteorological datasets.
• To develop and apply a novel machine learning (ML)-based framework for unbiasing and ensembling climate data to enhance crop modeling accuracy.

Main Methods:

• Utilized five gridded meteorological datasets (CFS, ERA5-Land, CHIRPS, IMERG, PERSIANN-CDR) for precipitation, temperature, and solar radiation.
• Applied a clustering-unbiasing-ensembling framework with four ML algorithms, including Light Gradient Boosting Machine and Random Forest.
• Implemented a TOTAL scenario combining all corrected variables and validated results using bootstrapping.

Main Results:

• Light Gradient Boosting Machine and Random Forest significantly improved precipitation and minimum temperature data accuracy (Nash-Sutcliffe Efficiency increased from 0.11-0.17 to 0.47 for precipitation, and 0.56-0.88 to 0.89-0.96 for temperature/solar radiation).
• The ML-based TOTAL scenario, correcting all variables, enhanced yield and water stress simulations in approximately 60% of cases.
• ML-based unbiasing-ensembling outperformed traditional methods, demonstrating reliability and transferability.

Conclusions:

• Advanced ML-based unbiasing-ensembling frameworks are valuable for improving climate data reliability in crop modeling.
• These methods offer practical solutions to support dryland agricultural systems facing climate extremes and data scarcity.
• The findings provide guidance for enhancing food security in vulnerable regions through improved climate data utilization.