Gül (2026) HydroSPI-ARX for explainable multi-scale drought early warning over Türkiye from ERA5 and ERA5-Land reanalyses

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Environmental Research Letters
• Year: 2026
• Date: 2026-11-03
• Authors: Ertuğrul Gül
• DOI: 10.1088/1748-9326/aebcff

Research Groups

Not specified in the abstract.

Short Summary

This study introduces HydroSPI-ARX, an autoregressive and explainable machine-learning framework for leakage-controlled drought early warning over Türkiye, demonstrating significant skill gains over baseline models, particularly at a six-month lead time, by leveraging the multi-scale memory of the Standardized Precipitation Index (SPI).

Objective

• To develop and evaluate HydroSPI-ARX, an autoregressive and explainable machine-learning framework for leakage-controlled, lead-resolved drought early warning over Türkiye, specifically targeting the Standardized Precipitation Index (SPI) using hydroclimatic fields.

Study Configuration

• Spatial Scale: Türkiye (country-wide).
• Temporal Scale: Drought forecasting at lead times of one, three, six, and twelve months. Drought frequency analysis period: 1981 to 2025.

Methodology and Data

• Models used: HydroSPI-ARX (autoregressive and explainable framework), Naive persistence, Per-cell linear autoregression, Mann-Kendall analysis.
• Data sources: ERA5 and ERA5-Land hydroclimatic fields (reanalysis), Station-derived labels (for verification).

Main Results

• The HydroSPI-ARX framework achieves macro-F1 scores for moderate-or-worse drought warnings of 0.853 at a one-month lead, 0.794 at three months, and 0.709 at six months, decreasing to 0.486 at twelve months.
• The framework outperforms naive persistence by 0.038 macro-F1 at three months and 0.066 at six months.
• Against a per-cell linear autoregression, a gain of 0.043 macro-F1 is observed only at six months (95% confidence interval: 0.010 to 0.078).
• Exogenous hydroclimatic fields contribute minimally to skill, adding 0.007 macro-F1 at one month and nothing beyond.
• Skill is primarily derived from the multi-scale memory of the Standardized Precipitation Index (SPI).
• Verification against station-derived labels shows lower agreement (Cohen's kappa: 0.30 at one month, 0.26 at three months, 0.21 at six months), but the ranking of forecasts is preserved.
• ERA5 and ERA5-Land reanalysis datasets are found to be equivalent within a 0.02 macro-F1 margin.
• A Mann-Kendall analysis indicates a rising three-month drought frequency over Türkiye from 1981 to 2025.

Contributions

• Introduction of HydroSPI-ARX, a novel leakage-controlled, autoregressive, and explainable framework for drought early warning.
• Demonstration of significant skill gains over traditional baselines (naive persistence, linear autoregression) for drought forecasting, particularly at a six-month horizon, by effectively leveraging the multi-scale memory of the SPI.
• Rigorous evaluation methodology controlling for temporal leakage and separating out the inherent memory of the drought index.
• Verification against independent station data, confirming the robustness of forecast rankings despite lower absolute agreement with reanalysis-derived targets.
• Assessment of reanalysis data equivalence (ERA5 vs. ERA5-Land) and the impact of climate zone information on model performance.

Funding

Not specified in the abstract.

Citation

@article{Gül2026HydroSPIARX,
  author = {Gül, Ertuğrul},
  title = {HydroSPI-ARX for explainable multi-scale drought early warning over Türkiye from ERA5 and ERA5-Land reanalyses},
  journal = {Environmental Research Letters},
  year = {2026},
  doi = {10.1088/1748-9326/aebcff},
  url = {https://doi.org/10.1088/1748-9326/aebcff}
}