Talebi et al. (2025) Advanced Hybrid Machine Learning for Precise Short-Term Drought Prediction: A Comparative Study of SPI and SPEI Indices in Iran's Arid and Semi-Arid Regions

Identification

• Journal: Pure and Applied Geophysics
• Year: 2025
• Date: 2025-11-28
• Authors: Hamed Talebi, Hatice Çıtakoğlu, Saeed Samadianfard, Aykut Erol
• DOI: 10.1007/s00024-025-03876-y

Research Groups

• Department of Water Engineering, University of Tabriz, Tabriz, Iran
• Department of Civil Engineering, Faculty of Engineering, Erciyes University, Melikgazi, Kayseri, Türkiye
• Water Sciences and Hydroinformatics Research Center, Khazar University, Baku, Azerbaijan
• Department of Environmental Engineering, İzmir Institute of Technology, Urla, Izmir, Türkiye

Short Summary

This study developed and compared twelve hybrid machine learning models for precise short-term drought prediction using SPI and SPEI indices in Iran's arid and semi-arid regions. It found that Tuned Q-factor Wavelet Transform (TQWT)-based models excelled in 1-month forecasts, while Empirical Wavelet Transform (EWT)-Adaptive Neuro-Fuzzy Inference System (ANFIS) was most robust for 3- and 6-month predictions.

Objective

• To systematically evaluate and identify the most reliable and robust hybrid machine learning framework for short- to medium-term meteorological drought prediction using Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) in Iran's arid and semi-arid regions.

Study Configuration

• Spatial Scale: Four synoptic meteorological stations in Iran (Tabriz, Shiraz, Kerman, Yazd) representing semi-arid and arid climate conditions across an area of approximately 1,648,195 square kilometers.
• Temporal Scale: Daily meteorological data from 1990 to 2022 (33 years) used to predict drought indices at 1-month, 3-month, and 6-month aggregation periods.

Methodology and Data

• Models used:
  • Signal Decomposition: Tuned Q-factor Wavelet Transform (TQWT), Variational Mode Decomposition (VMD), Empirical Mode Decomposition (EMD), Empirical Wavelet Transform (EWT).
  • Prediction Algorithms: Gaussian Process Regression (GPR), Support Vector Machines (SVM), Adaptive Neuro-Fuzzy Inference System (ANFIS).
  • Hybrid Models: Twelve combinations of decomposition and prediction algorithms (e.g., TQWT-GPR, VMD-SVM, EWT-ANFIS).
• Data sources:
  • Daily precipitation (mm), daily maximum temperature (°C), daily minimum temperature (°C), and daily mean temperature (°C) from the Iranian Meteorological Organization (IRIMO).
  • Data collected from 1990 to 2022 at four stations: Tabriz, Shiraz, Kerman, and Yazd.
  • Derived data: Monthly SPI and SPEI indices calculated at 1-month, 3-month, and 6-month aggregation periods.
  • Input variables for models included lagged drought index values (t-1, t-2, t-3).
  • Dataset split: 75% for training, 25% for testing.

Main Results

• TQWT-based models (TQWT-GPR and TQWT-ANFIS) demonstrated the highest predictive accuracy for 1-month drought forecasts. TQWT-ANFIS was best at Tabriz, while TQWT-GPR excelled at Shiraz, Kerman, and Yazd (R² ≈ 0.996–0.997; RMSE ≈ 0.05–0.07).
• For 3- and 6-month temporal evaluations, the EWT-ANFIS model consistently showed the best performance across all stations, achieving low error metrics (RMSE ≈ 0.01–0.03) and nearly perfect goodness-of-fit values (R² and NSE ≈ 0.999).
• Model performance generally improved with increasing time scale (from 1 to 6 months), characterized by declining error measures (MSE, RMSE, MAE) and increasing goodness-of-fit (R², NSE) towards 0.999.
• The correlation between SPI and SPEI strengthened with longer time scales across all stations, with semi-arid regions generally exhibiting higher R² values than arid ones, particularly at shorter time scales.
• Warming trends, especially in minimum temperatures, were observed to amplify evapotranspiration, thereby intensifying drought severity across the study region.

Contributions

• This study presents the first comparative analysis of SPI and SPEI prediction accuracy using hybrid decomposition-machine learning methods across Iran's distinct arid and semi-arid climate zones.
• It introduces and validates superior hybrid models (TQWT-GPR for short-term and EWT-ANFIS for medium-term predictions) that outperform existing benchmarks, often with fewer input variables.
• The research provides a robust and highly accurate predictive framework for operational drought monitoring and early warning systems, particularly valuable for water resource management in water-scarce regions.
• It highlights the critical role of SPEI in arid environments by accounting for temperature-driven evapotranspiration, offering a more comprehensive drought assessment.

Funding

No funding source.

Citation

@article{Talebi2025Advanced,
  author = {Talebi, Hamed and Çıtakoğlu, Hatice and Samadianfard, Saeed and Erol, Aykut},
  title = {Advanced Hybrid Machine Learning for Precise Short-Term Drought Prediction: A Comparative Study of SPI and SPEI Indices in Iran's Arid and Semi-Arid Regions},
  journal = {Pure and Applied Geophysics},
  year = {2025},
  doi = {10.1007/s00024-025-03876-y},
  url = {https://doi.org/10.1007/s00024-025-03876-y}
}