Ilyas et al. (2025) Modeling meteorological drought across scales with regional and global climate indicators

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-11-12
• Authors: Muhammad Ilyas, Rizwan Niaz, Hefa Cheng, Luca Di Persio, Maysaa Elmahi Abd Elwahab, Ali Danandeh Mehr, Ijaz Hussain
• DOI: 10.1007/s00704-025-05860-x

Research Groups

• Department of Statistics, Quaid-i-Azam University, Islamabad, Pakistan
• School of Energy and Environmental Science, Yunnan Normal University, Kunming, Yunnan, China
• College of Urban and Environmental Sciences, Peking University, Beijing, China
• Department of Computer Science, College of Mathematics, University of Verona, Verona, Italy
• Department of Mathematical Sciences, College of Science, Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia
• Department of Civil Engineering, Antalya Bilim University, Antalya, Türkiye

Short Summary

This study introduces a novel Hybrid Convolutional Bi-Kernel Ensemble (HCBKE) model for multiscale meteorological drought prediction using the standardized precipitation index (SPI) at 3, 6, and 12-month timescales. Evaluated in Ankara province, Turkey, the HCBKE model consistently outperformed traditional and deep learning models, demonstrating superior accuracy and robustness for operational drought monitoring.

Objective

• To develop and evaluate a novel Hybrid Convolutional Bi-Kernel Ensemble (HCBKE) model for accurate and timely multiscale meteorological drought prediction using the Standardized Precipitation Index (SPI) at 3, 6, and 12-month timescales, incorporating local climate indices and global teleconnection patterns.

Study Configuration

• Spatial Scale: Ankara Province, Turkey (approximately 39.875° N, 32.8333° E), spanning 24,521 square kilometers, at an average elevation of 850 meters above sea level. Six meteorological stations were selected: Beypazari, Esenboga, Kecioren, Kizilcahamam, Nallihan, and Polatli.
• Temporal Scale: A 52-year monthly time series from January 1971 to December 2022. Drought prediction was performed at 3, 6, and 12-month timescales (SPI-3, SPI-6, SPI-12).

Methodology and Data

• Models used:
  • Proposed: Hybrid Convolutional Bi-Kernel Ensemble (HCBKE), which synergistically integrates Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and kernel-based k-Nearest Neighbors (KNN).
  • Baseline/Comparative: Random Forest (RF), k-Nearest Neighbors (KNN), Gaussian Process Regression (GPR), Feedforward Neural Network (FNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM).
• Data sources:
  • Precipitation data: Turkish State Meteorological Service (TMD).
  • Temperature, soil moisture, relative humidity, wind speed, and surface pressure data: ERA5 reanalysis dataset (Copernicus Climate Data Store).
  • Global climate indices (North Atlantic Oscillation (NAO), El Niño-Southern Oscillation (ENSO), Mediterranean Oscillation Index (MOI)): NOAA and the University of East Anglia.
  • Input features: Six local standardized climatic condition indices (Precipitation Condition Index (PCI), Temperature Condition Index (TCI), Soil Moisture Condition Index (SMCI), Relative Humidity Condition Index (RHCI), Wind Speed Condition Index (WSCI), Surface Pressure Condition Index (SPCI)) and three global climate indices (NAO, ENSO, MOI).

Main Results

• The HCBKE model consistently demonstrated superior predictive capability across all SPI timescales and stations, outperforming traditional and deep learning models.
• For SPI-12 predictions at Polatlı, HCBKE achieved a maximum coefficient of determination (R²) of 0.788, a Nash-Sutcliffe Efficiency (NSE) of 0.788, and a minimum Root Mean Squared Error (RMSE) of 0.487.
• At SPI-3, despite higher variability, the model maintained moderate accuracy with R² values ranging from 0.363 to 0.448.
• The proposed model showed statistically significant improvements (p < 0.01) in Mean Absolute Error (MAE) across all timescales compared to baseline models, as confirmed by the Wilcoxon signed-rank test.
• The Mean Bias Error (MBE) remained statistically indistinguishable from zero (absolute value < 0.02) across all stations, indicating unbiased predictions.
• HCBKE reduced RMSE by 34% ± 5% over the best classical model (Random Forest) at SPI-3 and by 47% ± 7% over the best deep model (uni-directional LSTM) at SPI-12.
• Multicollinearity assessment using Variance Inflation Factor (VIF) identified high VIF values for TCI, SMCI, and RHCI; excluding SMCI and RHCI substantially reduced VIFs for remaining variables, ensuring a well-conditioned predictor set.

Contributions

• Introduced a novel Hybrid Convolutional Bi-Kernel Ensemble (HCBKE) model that synergistically integrates CNN, BiLSTM, and kernel-based KNN for robust multiscale meteorological drought prediction.
• Demonstrated superior and statistically significant predictive performance compared to a range of traditional and deep learning models across short (SPI-3), medium (SPI-6), and long-term (SPI-12) drought conditions.
• Addressed challenges of scale transferability, multicollinearity, and bias in drought forecasting through its unique architecture and a rigorous feature selection process.
• Provided a computationally efficient framework (approximately 1.5 minutes per station and horizon on an Intel i5 / 4 GB RAM machine), making it suitable for near-real-time early-warning applications.
• Achieved physically interpretable predictions by combining standardized local condition indices and global teleconnection patterns, with CNN filters learning to privilege PCI anomalies at shorter lags for SPI-3 and BiLSTM gates assigning higher attention to NAO phase shifts for SPI-12.

Funding

• Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2025R913), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Citation

@article{Ilyas2025Modeling,
  author = {Ilyas, Muhammad and Niaz, Rizwan and Cheng, Hefa and Persio, Luca Di and Elwahab, Maysaa Elmahi Abd and Mehr, Ali Danandeh and Hussain, Ijaz},
  title = {Modeling meteorological drought across scales with regional and global climate indicators},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05860-x},
  url = {https://doi.org/10.1007/s00704-025-05860-x}
}