Castaldo et al. (2026) Evaluating SEAS5 ECMWF seasonal forecasts for Sicily Mediterranean Island: Retrospective analysis and bias correction

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-02-24
• Authors: Francesco Castaldo, Antonio Francipane, Dario Treppiedi, Leonardo Noto
• DOI: 10.1016/j.ejrh.2026.103271

Research Groups

• Department of Engineering, University of Palermo, Palermo, Italy
• University School for Advanced Studies IUSS Pavia, Pavia, Italy

Short Summary

This study evaluates ECMWF SEAS5 seasonal forecasts for temperature and precipitation over Sicily, comparing traditional and Artificial Neural Network (ANN) bias correction methods. It finds that raw forecasts have systematic biases, and the ANN with Individual Member Separated Monthly (IMSM) correction significantly improves forecast accuracy, especially for precipitation, reducing Root Mean Square Error (RMSE) by up to 45 %.

Objective

• To assess the predictive capability of ECMWF SEAS5 seasonal forecasts for temperature and precipitation over Sicily and compare various bias correction techniques, including an Artificial Neural Network (ANN) approach, to improve forecast accuracy and reliability.

Study Configuration

• Spatial Scale: Sicily Island, Italy, focusing on four river basins (Piana degli Albanesi, Rosamarina, Poma, Scanzano) feeding artificial reservoirs near Palermo. The ECMWF SEAS5 forecast resolution is 1° × 1°, approximately 111 km × 88 km at the study latitude.
• Temporal Scale:
  • Observed data: April 1995 to June 2023 (Piana degli Albanesi, Poma, Scanzano basins) and February 2002 to June 2023 (Rosamarina basin).
  • Forecasts: Seasonal forecasts (SEAS5) with lead times (LT) from 0 to 6 months. Retrospective forecasts (Hindcasts, SHs) from 1995 to December 2016 (25 ensemble members) and Seasonal Real-Time forecasts (SRTs) from 2017 onwards (51 ensemble members).

Methodology and Data

• Models used:
  • ECMWF SEAS5 Seasonal Forecast system (numerical simulations of atmosphere, ocean, and land interactions).
  • Bias Correction Methods: Quantile Mapping (QM), Linear Regression (LR), Mean Bias Correction (MBC), Mean and Variance Adjustment (MVA), Artificial Neural Network (ANN).
  • ANN architecture: Feed-forward neural network with a single hidden layer (25 neurons for temperature, 23 for precipitation).
  • Bias Correction Approaches: Full Ensemble Median (FEM), Separated Monthly Median (SMM), Individual Member Separated Monthly (IMSM).
  • Performance Metrics: Mean Bias (MB), Root Mean Square Error (RMSE), Correlation Coefficient (CC), Kling-Gupta Efficiency (KGE), Continuous Ranked Probability Score (CRPS).
• Data sources:
  • Observed data: Mean monthly air temperature and monthly precipitation, spatially averaged across basins, provided by the River Basin Authority of the Sicily Region (AdB).
  • Forecast data: SEAS5 system of ECMWF, accessible via the Copernicus Climate Data Store.
  • Reanalysis data for SHs initialization: ERA-Interim.
  • Operational analyses for SRTs initialization: ECMWF operational analyses.
  • Digital Elevation Model (DEM): STR90 DEM (used by ECMWF for grid cell land surface characteristics).

Main Results

• Raw seasonal forecasts systematically overestimate low temperatures (e.g., Rosamarina basin mean bias up to 1.86 °C) and underestimate high precipitation (e.g., Piana degli Albanesi basin mean bias down to -45 mm, with forecasts never exceeding approximately 160 mm).
• The Artificial Neural Network (ANN) method combined with the Individual Member Separated Monthly (IMSM) approach achieved the best performance among all bias correction techniques.
• For precipitation, the ANN-IMSM method substantially improved forecast accuracy, reducing RMSE by up to 45 % (e.g., Piana degli Albanesi basin with ANN SRTs).
• Probabilistic forecast skill, measured by the Continuous Ranked Probability Score (CRPS), was significantly enhanced, with improvements up to 80 % for precipitation (e.g., November 2018, Piana degli Albanesi basin).
• The ANN-IMSM method effectively corrected both mean bias and error variability for precipitation while preserving ensemble variability.
• Using only Seasonal Real-Time forecasts (SRTs) for ANN training generally led to lower RMSE values compared to using both Hindcasts (SHs) and SRTs, attributed to greater data homogeneity.
• Forecast skill for precipitation depends on lead time and seasonal dynamics, with the lowest relative errors observed in November-February and the highest in August (e.g., Rosamarina basin normalized RMSE greater than 2 in August).

Contributions

• Provided a comprehensive comparison of traditional and advanced (ANN-based) bias correction techniques for seasonal forecasts in the vulnerable Mediterranean region (Sicily).
• Demonstrated the superior performance of an Artificial Neural Network (ANN) combined with an Individual Member Separated Monthly (IMSM) correction strategy for improving both deterministic and probabilistic seasonal forecasts of temperature and precipitation.
• Highlighted the critical importance of preserving ensemble variability through the IMSM approach for hydrological applications.
• Quantified the benefits of using more homogeneous Seasonal Real-Time forecasts (SRTs) for training bias correction models.
• Offered actionable insights for proactive water resource management and climate adaptation in semi-arid regions characterized by high intra-annual precipitation variability.

Funding

The authors gratefully acknowledge the Basin Authority of the Sicilian Region for providing the observed dataset. No specific funding projects, programs, or reference codes were listed.

Citation

@article{Castaldo2026Evaluating,
  author = {Castaldo, Francesco and Francipane, Antonio and Treppiedi, Dario and Noto, Leonardo},
  title = {Evaluating SEAS5 ECMWF seasonal forecasts for Sicily Mediterranean Island: Retrospective analysis and bias correction},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103271},
  url = {https://doi.org/10.1016/j.ejrh.2026.103271}
}