Tuğrul (2026) A performance comparison of machine learning algorithms for drought forecasting based on SPEI in Norway

Identification

• Journal: Meteorology and Atmospheric Physics
• Year: 2026
• Date: 2026-02-24
• Authors: Türker Tuğrul
• DOI: 10.1007/s00703-026-01119-w

Research Groups

• Department of Civil Engineering, Faculty of Technology, Gazi University, Ankara, Türkiye

Short Summary

This study compares the performance of five machine learning algorithms for forecasting drought based on the Standardized Precipitation Evapotranspiration Index (SPEI) in Bodo, Oslo, and Tromso, Norway. The research identifies the most effective models and input structures for each region, with SVM-M04 demonstrating the best overall performance across Oslo and Tromso, and LSTM-M01 for Bodo.

Objective

• To assess and compare the effectiveness of five advanced machine learning algorithms (LSTM, xGboost, SVM, MLP, and RF) in predicting monthly SPEI values for drought forecasting in Bodo, Oslo, and Tromso, Norway.
• To identify the optimal combination of algorithm and input structure that yields the highest statistical accuracy for site-specific drought predictions under diverse climatic conditions.

Study Configuration

• Spatial Scale: Three distinct regions in Norway: Bodo (temperate oceanic climate), Oslo (humid continental climate), and Tromso (subarctic climate). SPEI data has a 0.5 degree spatial resolution.
• Temporal Scale: Monthly SPEI data spanning from 1901 to 2023 (123 years). SPEI-12 was selected as the target output variable, with SPEI at 3, 6, 9, and 12-month timescales used as input data.

Methodology and Data

• Models used: Extreme Gradient Boosting (xGboost), Long-short Term Memory Network (LSTM), Support Vector Machine (SVM), Multi-layer Perceptron (MLP), and Random Forest (RF). The cross-correlation method was used to determine six different input structures (M01-M06) for the models.
• Data sources: Standardized Precipitation Evapotranspiration Index (SPEI) data obtained from the Global SPEI database (https://spei.csic.es/spei_database). 70% of the data was used for training and 30% for testing.

Main Results

• For Bodo, the LSTMM01 model achieved the highest performance (R = 0.9439, NSE = 0.8816, KGE = 0.8576, PI = 0.2642, RMSE = 0.3437).
• For Oslo, the SVMM04 model demonstrated the best performance (R = 0.9467, NSE = 0.8954, KGE = 0.8555, PI = 0.2696, RMSE = 0.3231).
• For Tromso, the SVMM04 model also yielded superior results (R = 0.9456, NSE = 0.8860, KGE = 0.8823, PI = 0.1921, RMSE = 0.3373).
• Overall, the SVMM04 model exhibited the most successful performance metrics across all regions, particularly for Oslo and Tromso.
• The cross-correlation method proved effective in determining suitable model input structures, with M01, M02, and M04 generally showing effective performance.
• The LSTM model showed high variability in performance, yielding both the highest and lowest accuracy scores depending on the input structure.

Contributions

• Provides a comprehensive performance comparison of five advanced machine learning algorithms (LSTM, xGboost, SVM, MLP, RF) for drought forecasting in Norway, a region with unique climatic conditions where droughts can have positive effects.
• Identifies optimal machine learning algorithms and specific input data structures (derived using cross-correlation) tailored for different Norwegian regions (Bodo, Oslo, Tromso), enhancing site-specific prediction accuracy.
• Utilizes SPEI-12, focusing on long-term hydrological drought, which is crucial for reservoir management and long-term planning in Nordic regions.
• Offers a robust decision-support tool for decision-makers to improve water resource management, agricultural planning, and ecological conservation in the study areas.

Funding

This study received no external funding.

Citation

@article{Tuğrul2026performance,
  author = {Tuğrul, Türker},
  title = {A performance comparison of machine learning algorithms for drought forecasting based on SPEI in Norway},
  journal = {Meteorology and Atmospheric Physics},
  year = {2026},
  doi = {10.1007/s00703-026-01119-w},
  url = {https://doi.org/10.1007/s00703-026-01119-w}
}