Phenology-Aware Machine Learning for Wheat Yield Prediction under Climate Variability: Central Anatolia, Türkiye

Abstract

This study aims to develop a phenology-aware machine learning framework for accurately predicting wheat yields in Türkiye’s Central Anatolia Region. The research integrates provincial wheat yield data from the Turkish Statistical Institute (TurkStat) (2004-2023) with fourteen agro-climatic and soil parameters retrieved from the National Aeronautics and Space Administration’s Prediction of Worldwide Energy Resources (NASA POWER) platform (2003-2023). To enhance model sensitivity, all variables were segmented into five key phenological stages of wheat growth, and for each stage, the minimum, maximum, and mean values were calculated. Three classical machine learning algorithms-Gradient Boosting (GB), Random Forest (RF), and Multilayer Perceptron (MLP)-were implemented using Python (Scikit-learn and TensorFlow libraries) under a “global training-local testing” strategy. The results show that GB consistently achieved the highest predictive accuracy across all provinces, with R2 values ranging from 0.96 to 0.99, mean absolute error (MAE) between 3.6 and 6.8 kg da-1, and root mean square error (RMSE) below 7.1 kg da-1. The RF model performed slightly lower (R2= 0.81-0.90) yet remained robust in most regions. In contrast, the global MLP model exhibited heterogeneous performance, particularly in Karaman Province, where non-climatic management factors dominate (R2= -1.25; MAE ? 26 kg da-1). When retrained with local data, the MLP model’s accuracy improved substantially, raising R2 to 0.79 and reducing MAE to approximately 10-15 kg da-1. These findings confirm that integrating phenological segmentation within ensemble learning approaches-particularly Gradient Boosting-substantially enhances wheat yield forecasting performance. The study highlights the importance of local calibration to capture irrigation and management effects and provides a robust methodological foundation for developing climate-resilient agricultural decision-support systems.

This study aims to develop a phenology-aware machine learning framework for accurately predicting wheat yields in Türkiye’s Central Anatolia Region. The research integrates provincial wheat yield data from the Turkish Statistical Institute (TurkStat) (2004-2023) with fourteen agro-climatic and soil parameters retrieved from the National Aeronautics and Space Administration’s Prediction of Worldwide Energy Resources (NASA POWER) platform (2003-2023). To enhance model sensitivity, all variables were segmented into five key phenological stages of wheat growth, and for each stage, the minimum, maximum, and mean values were calculated. Three classical machine learning algorithms-Gradient Boosting (GB), Random Forest (RF), and Multilayer Perceptron (MLP)-were implemented using Python (Scikit-learn and TensorFlow libraries) under a “global training-local testing” strategy. The results show that GB consistently achieved the highest predictive accuracy across all provinces, with R2 values ranging from 0.96 to 0.99, mean absolute error (MAE) between 3.6 and 6.8 kg da-1, and root mean square error (RMSE) below 7.1 kg da-1. The RF model performed slightly lower (R2= 0.81-0.90) yet remained robust in most regions. In contrast, the global MLP model exhibited heterogeneous performance, particularly in Karaman Province, where non-climatic management factors dominate (R2= -1.25; MAE ? 26 kg da-1). When retrained with local data, the MLP model’s accuracy improved substantially, raising R2 to 0.79 and reducing MAE to approximately 10-15 kg da-1. These findings confirm that integrating phenological segmentation within ensemble learning approaches-particularly Gradient Boosting-substantially enhances wheat yield forecasting performance. The study highlights the importance of local calibration to capture irrigation and management effects and provides a robust methodological foundation for developing climate-resilient agricultural decision-support systems.