Guzzon et al. (2025) Improving extreme precipitation forecasts in Catalonia (NE Iberian Peninsula) using analog methods: A comparison with the GFS model

Identification

• Journal: Weather and Climate Extremes
• Year: 2025
• Date: 2025-11-22
• Authors: Carlo Guzzon, Raül Marcos-Matamoros, María Carmen Llasat, Montserrat Llasat-Botija
• DOI: 10.1016/j.wace.2025.100839

Research Groups

• Department of Applied Physics, Universitat de Barcelona
• Earth Sciences Department, Barcelona Supercomputing Center - CNS
• UBICS, Universitat de Barcelona
• Water Research Institute, Universitat de Barcelona

Short Summary

This study evaluates novel analog-based methods (AMs) to enhance 24-hour extreme precipitation forecasts in Catalonia, aiming to support flood risk management. The findings demonstrate that AMs integrating Seasonal Standardization and the Perfect Prognosis framework significantly improve forecasts compared to the operational Global Forecast System (GFS), particularly in reproducing the intensity and spatial distribution of extreme events.

Objective

• To evaluate the potential of analog-based methods (AMs) to enhance 24-hour precipitation forecasts for Catalonia (northeastern Iberian Peninsula) with the broader objective of supporting flood risk management and early warning systems.
• To investigate whether AMs can improve the timing, location, and intensity of 24-hour extreme precipitation forecasts over Catalonia compared with operational GFS output, and whether they can provide operationally useful probabilistic guidance for flash flood early warning.

Study Configuration

• Spatial Scale:
  • Analog search domain: Western Mediterranean (D09), spanning from 17° W to 9° E longitude and 31° N to 48° N latitude.
  • Precipitation verification domain: Catalonia region, bounded by 0° W to 3.5° E longitude and 40° N to 43° N latitude.
  • GFS resolution: Approximately 13 km (0.25° for GFS-24h, 1° x 1° for GFS-ANL).
  • ERA5 resolution: 0.25° latitude–longitude grid.
• Temporal Scale:
  • Forecast lead time: 24 hours.
  • Precipitation Days pool: All days with recorded precipitation from 2004 to 2020 (3746 days).
  • INUNGAMA Days pool: 230 historical flood events in Catalonia from 1940 to 2020.
  • ERA5 Reanalysis data: 1940 to present (used 1940-2023 for WT-PP-S analog search, 1991-2020 for climatology).
  • GFS 0.25 Degree Forecast: 2015 to 2020.
  • GFS Analysis (GFS-ANL Grid 3): March 2004 to April 2020.

Methodology and Data

• Models used:
  • Analog-based methods (AMs):
    • No-WT AM (without weather type classification)
    • WT AM (with weather type classification)
    • WT-S AM (with weather types and Seasonal Standardization)
    • WT-PP-S AM (with weather types, Seasonal Standardization, and Perfect Prognosis assumption)
  • Global Forecast System (GFS):
    • GFS-24h: Operational GFS 0.25° forecast (00 UTC initialization).
    • GFS-6h: GFS Analysis (1° x 1°), aggregated from four daily 6-hour forecasts (used as a benchmark).
• Data sources:
  • ERA5 Reanalysis (ECMWF): Hourly accumulated precipitation and geopotential height fields at 500 hPa and 1000 hPa (0.25° resolution). Used as ground truth for verification and for the analog search in WT-PP-S.
  • Global Forecast System (GFS) data (U.S. National Weather Service):
    • GFS 0.25 Degree Forecast: Operational model outputs (0.25° resolution, 2015-2020).
    • GFS Analysis (GFS-ANL Grid 3): Historical outputs (1° x 1° resolution, March 2004-April 2020). Used for analog search (geopotential height) and to derive GFS-6h precipitation.
  • INUNGAMA database: Impact-based archive of flood events in Catalonia (1900-2020).
  • Predictors for AMs: Geopotential height fields at 500 hPa (Z500) and 1000 hPa (Z1000).
  • Analog ranking metrics: Combined Euclidean distance and Pearson spatial correlation distance.
  • Verification metrics: Mean Error (ME), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Probability Of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI), Pearson correlation coefficient (r).

Main Results

• General Precipitation Events (Precipitation Days Pool):
  • All AMs reduce the underestimation of heavy precipitation compared to GFS-24h, with WT, WT-S, and WT-PP-S better reproducing the distribution of heavy rainfall (≥ 30 mm/24 h).
  • WT-S and WT-PP-S achieve lower MAE and RMSE than GFS-24h, showing performance comparable to GFS-6h.
  • WT-PP-S outperforms GFS-24h in POD and CSI for 5 mm, 10 mm, and 30 mm thresholds, but GFS-6h generally maintains higher skill, especially in reducing false alarms.
  • WT-S and WT-PP-S consistently outperform GFS-24h in reproducing the spatial distribution of precipitation (Pearson spatial correlation coefficient, r), though GFS-6h remains superior.
• Extreme Precipitation Events (INUNGAMA Days Pool):
  • All forecasting methods tend to underestimate maximum precipitation peaks exceeding 40 mm/24 h.
  • WT-S and WT-PP-S models show a marked improvement over GFS-24h in reproducing the distribution of higher precipitation levels, reducing the underestimation bias.
  • WT-S and WT-PP-S exhibit reduced negative bias in ME, with WT-PP-S showing ME values closest to zero.
  • WT-S and WT-PP-S consistently outperform GFS-24h in MAE and RMSE.
  • WT-PP-S substantially improves POD and CSI compared to GFS-24h across 20 mm, 40 mm, and 80 mm thresholds, demonstrating competitive performance for extreme events.
  • WT-S and WT-PP-S achieve better spatial correlation (r) than GFS-24h.
• Sensitivity Analysis: Averaging up to five analogs results in largely stable RMSE and Pearson correlation coefficients, with only minor degradation when including all 10 analogs, supporting the use of a 10-analog ensemble mean for robust forecasts.

Contributions

• Introduces a novel analog-based forecasting framework that integrates Weather-Type classification, Seasonal Standardization, and the Perfect Prognosis assumption.
• Demonstrates that these enhanced analog methods (WT-S and WT-PP-S) significantly improve 24-hour extreme precipitation forecasts in Catalonia compared to the operational GFS model.
• Shows that the Perfect Prognosis approach, by leveraging an extended historical analog pool (83 years from ERA5), enhances the ability to identify suitable analogs for rare and intense events, improving skill for extreme precipitation.
• Highlights the operational potential of these efficient, data-driven methods as complements to numerical weather prediction models for flash-flood forecasting and impact-based risk management, offering improved skill in capturing localized extremes and generating probabilistic information.

Funding

• Project PLEC2022-009403, Flood2Now – Improvement of early warning systems for flood risk based on past information and citizen science data, funded by MCIN/AEI/10.13039-/501100011033 and the European Union NextGenerationEU/PRTR.
• Inund-IA, funded by the Agència Catalana de l’Aigua de la Generalitat de Catalunya.

Citation

@article{Guzzon2025Improving,
  author = {Guzzon, Carlo and Marcos-Matamoros, Raül and Llasat, María Carmen and Llasat-Botija, Montserrat},
  title = {Improving extreme precipitation forecasts in Catalonia (NE Iberian Peninsula) using analog methods: A comparison with the GFS model},
  journal = {Weather and Climate Extremes},
  year = {2025},
  doi = {10.1016/j.wace.2025.100839},
  url = {https://doi.org/10.1016/j.wace.2025.100839}
}