Estimation of Winter Wheat Yield Using Multi-Source Remote Sensing Data and Machine Learning in an Arid and Semi-Arid Region

Introduction:

Reliable estimation of crop yield is crucial for ensuring food security and sustainable agricultural management, particularly in arid and semi-arid regions with sparse monitoring networks.  Although remote sensing and machine learning have advanced yield prediction, model performance depends strongly on the selection and integration of spectral, climatic, and crop-related information. Few studies have simultaneously integrated optical remote sensing, remotely sensed evapotranspiration, and climate data within a phenology-aware machine learning framework, particularly across spatially heterogeneous agricultural environments. This study addresses this gap by evaluating a multi-source, phenology-based framework for winter wheat yield estimation.

 Methods:

Irrigated winter wheat yield was estimated across seven counties in Fars Province, Iran, using Sentinel-2 vegetation indices, WaPOR actual evapotranspiration (ETa), AgERA5 meteorological data, and geographic information.  Predictors were extracted according to crop phenological stages, and three machine learning algorithms, Random Forest (RF), Extreme Gradient Boosting (XGB), and Elastic Net (EN), were evaluated.

Results:

The nonlinear algorithms (RF and XGB) consistently outperformed the linear EN, while the three-stage framework provided higher accuracy and robustness than the eight-stage framework.  RF achieved the best performance (NRMSE = 0.28; MAE = 0.76 t ha⁻¹). Geographic location was the most influential predictor, followed by MERIS Terrestrial Chlorophyll Index (MTCI) during flowering, dew point temperature during maturity, diurnal temperature range during flowering, cumulative rainfall during the vegetative stage, and ETa. Importantly, AgERA5 successfully replaced ground-based meteorological observations with only a marginal reduction in prediction accuracy.

Conclusions:

The proposed phenology-aware integration of multi-source remote sensing products with machine learning offers a practical and scalable framework for estimating winter wheat yields in data-scarce arid and semi-arid environments. The findings highlight the potential of satellite-derived datasets and artificial intelligence to support operational agricultural monitoring and precision agriculture where conventional observation networks are limited.