Anık et al. (2025) Investigating the contribution of decomposition techniques to machine learning accuracy in SPEI-based drought forecasting for multiple Köppen-Geiger climates

Identification

• Journal: Acta Geophysica
• Year: 2025
• Date: 2025-12-23
• Authors: Emirhan Mustafa Anık, Burçe Toğrul, Abdullah Akbaş, Murat Kankal
• DOI: 10.1007/s11600-025-01773-5

Research Groups

• Department of Civil Engineering, Bursa Uludağ University, Türkiye
• Department of Geography, Faculty of Arts and Science, Bursa Uludağ University, Türkiye

Short Summary

This study investigates the impact of various decomposition techniques on machine learning accuracy for SPEI-based drought forecasting across different Köppen-Geiger climates. The research found that decomposition methods significantly enhance prediction performance, with Variational Mode Decomposition (VMD) proving most effective, leading to Nash–Sutcliffe Efficiency (NSE) values consistently above 0.95 across all SPEI time scales.

Objective

• To evaluate the forecasting of droughts using a hybrid approach that combines various decomposition techniques (VMD, DWT, EMD, EEMD) and machine learning methods (MLP, KAN, RNN, BiLSTM, BiGRU, RF, GB, XGB, AB, M5P) for different climate regions.
• To assess the performance of four decomposition techniques: VMD, DWT, EMD, and EEMD.
• To evaluate the performance of ten machine learning methods, including network-based and tree-based approaches.
• To examine the effect of climate class on the accuracy of drought forecasting using station data from different Köppen-Geiger classifications.
• To apply the Kolmogorov-Arnold Network (KAN) method for the first time in drought forecasting.

Study Configuration

• Spatial Scale: Three meteorological stations in Türkiye (Bursa, Erzurum, Konya), representing three distinct Köppen-Geiger climate classifications (Csa: Temperate, dry hot summer; Dfb: Snow/continental, no dry season, warm summer; BSk: Arid, steppe, cold).
• Temporal Scale: Monthly precipitation and temperature data spanning 52 years (1969 to 2020). Drought predictions were made for SPEI time scales of 3, 6, and 12 months.

Methodology and Data

• Models used:
  • Decomposition Techniques: Variational Mode Decomposition (VMD), Discrete Wavelet Transform (DWT), Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD).
  • Machine Learning Methods:
    • Network-based: Multilayer Perceptrons (MLP), Kolmogorov-Arnold Networks (KAN), Recurrent Neural Networks (RNN), Bidirectional Long Short-Term Memory (BiLSTM), Bidirectional Gated Recurrent Unit (BiGRU).
    • Tree-based: Random Forest (RF), Gradient Boosting (GB), eXtreme Gradient Boosting (XGB), Adaptive Boosting (AB), M5Prime (M5P).
  • Drought Index: Standardised Precipitation Evapotranspiration Index (SPEI), calculated using the R programming language 'spei' package.
  • Performance Metrics: Nash–Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), Pearson Correlation Coefficient (R).
  • Visual Evaluation: Taylor diagrams, Standard Triangular Diagram (STD).
  • Explainable AI: Shapley Additive Explanations (SHAP) for feature contribution analysis.
• Data sources: Monthly precipitation and temperature data from the Turkish State Meteorological Service. Data homogeneity and sufficiency were evaluated using the standard normal homogeneity test, Pettitt test, and Buishand rank test.

Main Results

• Decomposition techniques significantly improved drought prediction accuracy across all climate types and SPEI timescales compared to raw data.
• Variational Mode Decomposition (VMD) was identified as the most effective decomposition technique, followed by DWT. EMD and EEMD generally underperformed compared to raw data models.
• For SPEI-3, NSE values, which were approximately 0.5 with raw data, increased to above 0.95 after VMD implementation.
• For SPEI-6, NSE values, approximately 0.7 with raw data, increased to above 0.95 with VMD.
• For SPEI-12, NSE values, approximately 0.9 with raw data, increased to above 0.95 with VMD.
• MLP, KAN, and M5P consistently demonstrated the highest performance among machine learning methods, achieving NSE values over 0.977 in test data when combined with VMD.
• Recurrent neural network-based models (RNN, BiLSTM, BiGRU) showed improved performance with increasing time scales, reaching accuracy levels comparable to MLP, KAN, and M5P for SPEI-6 and SPEI-12.
• The effect of climate class on model accuracy, which was evident when using raw data, largely disappeared after applying decomposition techniques, indicating enhanced model robustness.
• SHAP analysis revealed that low-frequency components (IMF1 and IMF2) and lagged features (especially t-1) were the most significant contributors to model performance.

Contributions

• Provided a comprehensive comparative analysis of four decomposition techniques and ten machine learning algorithms for SPEI-based drought forecasting across diverse Köppen-Geiger climate types.
• Introduced the Kolmogorov-Arnold Network (KAN) method to drought forecasting for the first time, demonstrating its high performance and interpretability through explicit mathematical relationships.
• Demonstrated that decomposition methods, particularly VMD, are crucial for significantly enhancing drought forecasting accuracy and making models robust across different climate regions, effectively mitigating the influence of climatic conditions.
• Offered insights into optimal lag and decomposition level selections for various SPEI timescales and stations, reducing model complexity and preventing overfitting.
• Utilized SHAP analysis to provide explainable insights into the contributions of different features (decomposition level and lag) to model performance.

Funding

• TUBITAK 2211A (for the PhD studies of the first author)
• Scientific Research Projects Unit of Bursa Uludağ University (Performance-Based Support Project, Project Code: FPDD-2025–2307)

Citation

@article{Anık2025Investigating,
  author = {Anık, Emirhan Mustafa and Toğrul, Burçe and Akbaş, Abdullah and Kankal, Murat},
  title = {Investigating the contribution of decomposition techniques to machine learning accuracy in SPEI-based drought forecasting for multiple Köppen-Geiger climates},
  journal = {Acta Geophysica},
  year = {2025},
  doi = {10.1007/s11600-025-01773-5},
  url = {https://doi.org/10.1007/s11600-025-01773-5}
}