Çıtakoğlu et al. (2025) Multiscale drought forecasting via temporal–spectral decomposition and machine learning integration

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-10-22
• Authors: Hatice Çıtakoğlu, Veysi Kartal, Emre Topçu, Erdinç İkincioğulları, Mehmet Güney
• DOI: 10.1007/s00704-025-05836-x

Research Groups

• Civil Engineering Department, Faculty of Engineering, Erciyes University, Kayseri, Türkiye
• Civil Engineering Department, Engineering Faculty, Siirt University, Siirt, Türkiye
• Civil Engineering Department, Faculty of Engineering and Architecture, Kafkas University, Kars, Türkiye
• Department of Civil Engineering, Bingol University, Bingol, Türkiye
• Graduate School of Natural and Applied Sciences, Civil Engineering of Graduate Program, Erciyes University, Kayseri, Türkiye

Short Summary

This study developed a novel multiscale drought forecasting framework by integrating temporal–spectral decomposition techniques with machine learning models to predict the Multivariate Standardized Drought Index (MSDI) at 1-, 3-, and 6-month time scales for the Sakarya region, Türkiye, finding the TQWT-GPR hybrid model to be the most accurate.

Objective

• To develop a novel multiscale drought forecasting framework that integrates temporal–spectral signal decomposition techniques with state-of-the-art supervised machine learning models to accurately anticipate future drought episodes using the Multivariate Standardized Drought Index (MSDI) at 1-, 3-, and 6-month accumulation intervals.

Study Configuration

• Spatial Scale: Sakarya province, Türkiye (between 29° 57’ and 30° 53’ east longitudes and 40° 17’ and 41° 13’ north latitudes). ERA5 dataset spatial resolution: 0.25° × 0.25°.
• Temporal Scale: January 1940 to December 2024 (85-year monthly time series). Drought time frames: 1, 3, and 6 months.

Methodology and Data

• Models used:
  • Drought Index: Multivariate Standardized Drought Index (MSDI) calculated using DroughtStats software.
  • Decomposition Methods (Pre-processing): Tunable Q-factor Wavelet Transform (TQWT), Maximal Overlap Discrete Wavelet Transform (MODWT), Variational Mode Decomposition (VMD).
  • Machine Learning Regression Models: Gaussian Process Regression (GPR), Support Vector Machine Regression (SVR), Bootstrap Aggregation (Bagging Regressor), Least Squares Boosting (LSBoost).
  • Input variables for ML models: lagged MSDI values (t-1, t-2, t-3).
  • Performance Metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Relative Root Mean Square Error (RRMSE), Coefficient of Determination (R²), Nash–Sutcliffe Efficiency (NSE), Kling–Gupta Efficiency (KGE), Overall Index (OI).
• Data sources: ERA5 reanalysis dataset (European Centre for Medium-Range Weather Forecasts - ECMWF). Variables: Volumetric Soil Layer 1 (0–7 cm surface layer) and precipitation (p). Monthly aggregates.

Main Results

• The hybrid TQWT-GPR model consistently achieved the best performance across all time scales for predicting MSDI.
• For the 1-month time scale (MSDI-1), TQWT-GPR yielded RRMSE = 8.08%, KGE = 0.998, and NSE = 0.997.
• For the 3-month time scale (MSDI-3), TQWT-GPR achieved RRMSE = 6.754%, KGE = 0.997, and NSE = 0.997.
• For the 6-month time scale (MSDI-6), TQWT-GPR demonstrated the highest accuracy with RRMSE = 3.70%, KGE = 0.995, and NSE = 0.999.
• The decomposition success of TQWT was found to be higher than that of VMD and MODWT.
• SVR-based hybrid models consistently exhibited lower performance compared to other models.

Contributions

• Proposes a novel multiscale drought forecasting framework integrating temporal–spectral signal decomposition with supervised machine learning models.
• Utilizes the Multivariate Standardized Drought Index (MSDI) as a core predictive variable, addressing the limitations of univariate indices.
• Evaluates and demonstrates the superior performance of the Tunable Q-factor Wavelet Transform (TQWT) as a pre-processing technique for drought signal decomposition.
• Identifies the TQWT-GPR hybrid model as the most effective for short- to long-term drought forecasting (1, 3, and 6 months), providing a robust and dependable framework.
• Contributes to the development of early warning systems and adaptive water resource management strategies under climate variability.

Funding

• Research Fund of the Erciyes University (Project FDK-2024-13905)

Citation

@article{Çıtakoğlu2025Multiscale,
  author = {Çıtakoğlu, Hatice and Kartal, Veysi and Topçu, Emre and İkincioğulları, Erdinç and Güney, Mehmet},
  title = {Multiscale drought forecasting via temporal–spectral decomposition and machine learning integration},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05836-x},
  url = {https://doi.org/10.1007/s00704-025-05836-x}
}