Tuğrul et al. (2025) Hybrid Wavelet–ML models for regional drought forecasting in Norway

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-11-04
• Authors: Türker Tuğrul, Sertaç Oruç, Jonathan P. Hall, Ali Ulvi Galip Şenocak, Mehmet Ali Hınıs
• DOI: 10.1038/s41598-025-22416-1

Research Groups

• Technology Faculty, Civil Engineering Department, Gazi University, Ankara, Türkiye
• The Center for Sámi Studies, UiT Norges Arktiske Universitet, Tromsø, Norway
• Faculty of Engineering and Natural Sciences, Department of Civil Engineering, Ankara Yıldırım Beyazıt University, Ankara, Türkiye
• The Arctic Youth Network and The Foundation for Law and International Affairs, Washington, USA
• Faculty of Engineering, Civil Engineering, Aksaray University, Aksaray, Türkiye

Short Summary

This study develops and evaluates hybrid wavelet-machine learning models for regional drought forecasting in Norway using the Effective Drought Index (EDI). The main finding is that Long Short-Term Memory (LSTM) networks enhanced by wavelet transformation (LSTMW) provide the best forecasts across the studied regions, significantly improving predictive accuracy.

Objective

• To develop and compare hybrid wavelet-machine learning models for forecasting the Effective Drought Index (EDI) in three climatically diverse Norwegian regions (Drammen, Hamar, and Lillehammer).
• To identify the most effective machine learning algorithm and input structure for regional drought prediction in Norway, addressing a gap in the literature regarding EDI application and geographical focus on high-latitude regions.

Study Configuration

• Spatial Scale: Three distinct regions in Norway: Drammen, Hamar, and Lillehammer.
• Temporal Scale: Monthly precipitation data from 1980 to 2025, used to calculate and forecast monthly Effective Drought Index (EDI).

Methodology and Data

• Models used: Support Vector Machine (SVM), Multi-layer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), Long Short-Term Memory network (LSTM), Categorical Boosting Algorithm (Catboost). These models were also hybridized with Discrete Wavelet Transform (DWT) (e.g., LSTMW, SVMW).
• Data sources: Monthly precipitation data from meteorological stations in Drammen, Hamar, and Lillehammer. The Effective Drought Index (EDI) was calculated from this precipitation data.

Main Results

• Wavelet Transformation (WT) consistently improved the performance metrics across all machine learning algorithms and nearly all input configurations.
• Long Short-Term Memory (LSTM) models, particularly when enhanced by WT (LSTMW), generally exhibited the best forecasting performance across the regions.
• For Drammen, the LSTMW-M04 model achieved the highest performance with a correlation coefficient (r) of 0.9765, Nash-Sutcliffe Efficiency (NSE) of 0.9510, Kling-Gupta Efficiency (KGE) of 0.8641, Performance Index (PI) of 0.3211, and Root Mean Square Error (RMSE) of 0.2207.
• For Hamar, LSTMW-M03 was the top performer (r = 0.9689, NSE = 0.9369, KGE = 0.8747, PI = 0.3433, RMSE = 0.2503).
• For Lillehammer, the SVMW-M03 model showed superior performance (r = 0.9635, NSE = 0.9258, KGE = 0.8868, PI = 0.3666, RMSE = 0.2716), distinguishing it from the other regions where LSTMW was dominant.
• Multi-layer Perceptron (MLP) models generally demonstrated the weakest performance, although WT improved their metrics from negative to positive values in some instances.
• Statistical significance tests (ANOVA and Kruskal-Wallis) indicated no significant difference between the mean predictions of the top-performing models and the observed test-set data (p-values > 0.05).

Contributions

• Introduces a novel application of hybrid Wavelet Transformation with a comprehensive suite of machine learning algorithms for regional drought forecasting in Norway.
• Represents the first extensive application of the Effective Drought Index (EDI) in a Norwegian context, particularly integrated with advanced machine learning and deep learning techniques for regionally specific predictions.
• Addresses a significant geographical gap in drought prediction literature by focusing on high-latitude regions of Norway (Drammen, Hamar, Lillehammer), which have been less studied despite increasing drought frequency.
• Provides a comparative evaluation of five advanced ML/DL algorithms, demonstrating their relative performance and calibration potential for Norway's diverse climatic and agricultural zones.
• Offers a practical and applicable framework for developing localized, anticipatory water management systems and informs future agricultural, energy, and municipal planning in climate-sensitive regions.

Funding

• Open access funding provided by UiT The Arctic University of Norway (including University Hospital of North Norway).
• No external funding was received for this research.

Citation

@article{Tuğrul2025Hybrid,
  author = {Tuğrul, Türker and Oruç, Sertaç and Hall, Jonathan P. and Şenocak, Ali Ulvi Galip and Hınıs, Mehmet Ali},
  title = {Hybrid Wavelet–ML models for regional drought forecasting in Norway},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-22416-1},
  url = {https://doi.org/10.1038/s41598-025-22416-1}
}