Evaluation and selection of subseasonal-to-seasonal (S2S) models for rainfall prediction over Ethiopia

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Numerous subseasonal-to-seasonal (S2S) models exist, yet their performance over tropical regions remains poorly understood, leaving forecasters with little guidance on which models best suit operational downscaling. This study evaluates 14 global models from the C3S and NMME initiatives for precipitation prediction over Ethiopia. We utilize Taylor diagrams and spatial bias assessments to categorize model performance across monthly, seasonal, and annual scales, using CHIRPS and ENACTS datasets as references. Our quantitative results reveal substantial inter-model variability, with correlation coefficients ranging from 0.46 to 0.92 and Root Mean Square Error (RMSE) values between 33.46 and 80.89 mm/month. The JMA, ECCC, and ECMWF models consistently emerge as top performers, maintaining high correlations (0.88–0.92) and the lowest RMSE values (33.46–35.84 mm/month). Seasonal analysis demonstrates that these three models effectively capture the spatial climatology of the main rainy season (JJAS) and the ONDJ season. However, the JJAS season exhibits the highest one-standard-deviation spread, indicating increased forecast uncertainty during peak rainfall periods. Conversely, the NMME ensemble tends to overestimate precipitation during the FMAM season. We find that while individual C3S models generally exhibit higher skill, the full NMME ensemble outperforms the C3S ensemble through beneficial error cancellation. Annual skill scores confirm that a parsimonious ensemble of JMA, ECCC, and ECMWF reduces spatial bias more effectively than larger aggregates. Consequently, we recommend this refined subset for operational services and downscaling at the Ethiopian Meteorological Institute (EMI). This selection optimizes predictive capability for Ethiopia’s tropical climate while reducing computational overhead and simplifying data management.

This publication has 13 references indexed in Scilit: