Güngüneş et al. (2025) Phenology-Aware Machine Learning for Wheat Yield Prediction under Climate Variability: Central Anatolia, Türkiye

Identification

• Journal: Türkiye Tarımsal Araştırmalar Dergisi
• Year: 2025
• Date: 2025-10-31
• Authors: Ramazan Güngüneş, Volkan Ateş, Taşkın Erol, Rojin Özek
• DOI: 10.19159/tutad.1740059

Research Groups

• Kırıkkale University, Kırıkkale Vocational School, Department of Electricity and Energy, Electrical Program, Kırıkkale, TÜRKİYE
• Tarsus University, Faculty of Engineering, Department of Computer Engineering, Tarsus, TÜRKİYE
• Kırıkkale University, Kırıkkale Vocational School, Department of Plant and Animal Production, Organic Farming Program, Kırıkkale, TÜRKİYE

Short Summary

This study developed a phenology-aware machine learning framework for wheat yield prediction in Türkiye's Central Anatolia Region, demonstrating that Gradient Boosting consistently achieved high accuracy (R² 0.96-0.99) by integrating phenological segmentation of agro-climatic and soil parameters. The research highlights the critical role of local calibration to account for management practices like irrigation.

Objective

• To develop and evaluate a phenology-aware machine learning framework for accurate wheat yield prediction in Türkiye's Central Anatolia Region under climate variability, integrating agro-climatic and soil parameters segmented by phenological stages.

Study Configuration

• Spatial Scale: Central Anatolia Region, Türkiye, with province-level evaluations across 12 provinces.
• Temporal Scale: Wheat yield data from 2004-2023 (20 years) and agro-climatic/soil parameters from 2003-2023 (21 years), segmented into five key phenological stages (Sowing and emergence, Tillering, Stem elongation and heading, Milk development, Harvesting).

Methodology and Data

• Models used: Gradient Boosting (GB), Random Forest (RF), Multilayer Perceptron (MLP). Implemented using Python (Scikit-learn and TensorFlow-Keras libraries).
• Data sources:
  • Provincial wheat yield data (2004-2023) from the Turkish Statistical Institute (TurkStat).
  • Fourteen agro-climatic and soil parameters (2003-2023) from the National Aeronautics and Space Administration’s Prediction of Worldwide Energy Resources (NASA POWER) platform.
  • Variables were segmented into five phenological stages, and minimum, maximum, and mean values were calculated for each stage.

Main Results

• The Gradient Boosting (GB) model consistently achieved the highest predictive accuracy across all provinces, with R² values ranging from 0.96 to 0.99.
• GB model's mean absolute error (MAE) was between 0.0036 kg m⁻² and 0.0068 kg m⁻², and root mean square error (RMSE) was below 0.0071 kg m⁻².
• The Random Forest (RF) model performed robustly but slightly lower, with R² values typically between 0.81 and 0.90.
• The Multilayer Perceptron (MLP) model exhibited heterogeneous performance, particularly poor in Karaman Province (R² = -1.25; MAE ≈ 0.026 kg m⁻²), where non-climatic management factors are dominant.
• Local retraining of the MLP model with local data substantially improved its accuracy, raising R² to 0.79 and reducing MAE to approximately 0.010-0.015 kg m⁻².
• Integrating phenological segmentation within ensemble learning approaches, especially Gradient Boosting, significantly enhances wheat yield forecasting performance.

Contributions

• Confirms that the integration of phenological segmentation and ensemble learning significantly enhances the temporal and spatial precision of wheat yield prediction models.
• Provides a methodological basis for bridging the gap between process-based crop models and artificial intelligence (AI) applications in agricultural forecasting.
• Highlights the critical importance of local calibration and management information (e.g., irrigation, fertilization) for improving model reliability and capturing regional heterogeneity.
• Establishes a robust methodological foundation for developing climate-resilient agricultural decision-support systems that integrate AI with on-ground agronomic management.

Funding

• This research received no external funding.

Citation

@article{Güngüneş2025PhenologyAware,
  author = {Güngüneş, Ramazan and Ateş, Volkan and Erol, Taşkın and Özek, Rojin},
  title = {Phenology-Aware Machine Learning for Wheat Yield Prediction under Climate Variability: Central Anatolia, Türkiye},
  journal = {Türkiye Tarımsal Araştırmalar Dergisi},
  year = {2025},
  doi = {10.19159/tutad.1740059},
  url = {https://doi.org/10.19159/tutad.1740059}
}