Evaluation of the Skill of Seasonal Rainfall and Temperature Forecasts From Global Prediction Models Over Ethiopia

Abstract

The skill of seasonal climate predictions from global forecasting models varies considerably across different regions and seasons. Evaluating the skill of these models in forecasting temperature and rainfall for various seasons is crucial for enhancing forecast accuracy and ensuring the efficient utilization of forecast information. This study analyzed the performance of the North American Multi-Model Ensemble (NNME) and Copernicus Climate Change Service (C3S) seasonal forecast models in predicting the June-September (JJAS) and February-May (FMAM) seasonal rainfall and temperature over Ethiopia. The seasonal forecast models were evaluated for the hindcast period of 1994–2016 using the Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) data as a reference for rainfall and the Climatic Research Unit (CRU) for temperature. The skill assessment is conducted using the recently developed Python-based Climate Predictability Tool (PyCPT). The model predictor domains used were tropical ocean, the Western Indian Ocean, Ethiopia itself, the Atlantic and Indian Oceans at forecast time leads of 2-months, 1-month and 0- month. The models’ skill assessment was performed using various skill scores such as Pearson correlation, Relative Operating Characteristics (ROC) and Ranked Probability Skill Score (RPSS). The results show that the NMME and C3S models show varying levels of the skill of forecasted seasonal rainfall. The better performing models with the Pearson correlation greater than 0.5, ROC below score exceeding 0.7, ROC above score exceeding 0.6 and RPSS greater than 20% are CanSIPS-IC3, ECMWF-SEAS5, DWD-GCFS2P1, CMCC-SPS3P5 and METEOFRANCE8. These models show higher ability in forecasting rainfall in the JJAS season over the Central, Northeastern, Northern, Northwestern and pocket areas in the Eastern portion of Ethiopia compared to other models like, GFDL-SPEAR, RSMAS-CCSM4, NASA- GEOSS2S and NCEPCFSv2. However, most models have low skill (Pearson correlation <0, ROC<0.5 and RPSS <0) to predict the rainfall in the JJAS season over the Southern, Southeastern, Southwestern and Western half of Ethiopia. It has been indicated that May and June initialized forecasts show better skill compared to the April initialized forecasts during the JJAS season. For effective agricultural and water management planning, the preferred choice is the May-initialized forecast. During the v | P a g e FMAM season, the CanSIPS-IC3, ECMWF-SEAS5, DWD-GCFS2P1, CMCC-SPS3P5 and METEOFRANCE8 models show higher skill (Pearson correlation between 0.5 & 0.7, ROC below between 0.6 & 0.8, ROC above between 0.6 & 0.7 and RPSS between 20 & 40%) in forecasting the seasonal rainfall in the Southeastern, Southern, Central, and Eastern Ethiopia compared to other models like, RSMAS-CCSM4, NASA-GEOSS2S and NCEP-CFSv2. But their skill is lower in the Western half of Ethiopia. The 1-month lead forecast (initialized in January) exhibits a better skill compared to the 2-month and 0-month lead time forecast. The evaluation of various predictor domains' effect on forecast skill shows that the tropical domain (180°W to 180°E, 30°S to 30°N) exhibit higher skill in the JJAS and FMAM seasonal rainfall forecasts for Ethiopia when compared to other domains. It has been noted that the skill of models in forecasting the below normal rainfall are higher than the above normal rainfall for both JJAS and FMAM rainfall seasons