VERIFICATION OF BELG RAINFALL FORECASTS OVER SOUTHERN ETHIOPIA USING COPERNICUS CLIMATE CHANGE SERVICE (C3S) MODELS

Abstract

Southern Ethiopia, characterized by diverse topography and a bimodal rainfall pattern, relies heavily on the Belg (February-May) season for agricultural productivity and livelihood support. However, the region faces challenges due to the high rainfall variability during this season, necessitating reliable seasonal forecast systems to cope with the challenges. This thesis examined the rainfall forecasting performance of five Copernicus Climate Change Service (C3S) models: ECMWF, Météo-France, DWD, ECCC, and JMA in Belg rainfall during the hindcast period spanning 1993 to 2016. The surface observation data and Enhancing National Climate Services (ENACTS) dataset were used as reference. Probabilistic and deterministic metrics were used as the performance measure. The performance is evaluated on monthly and seasonal timescales by considering a one-month lead time. In addition, the performance of the Ethiopian Meteorology Institute (EMI) probabilistic seasonal forecast was compared with best performing dynamical climate model forecasting. Results revealed varying performance patterns of the models across different regions and months, highlighting specific areas where some models demonstrated strong predictive capabilities and potential areas for improvement in forecasting climatological rainfall. All the models relatively perform better at the start of the season. The spatial analysis depicts that the performance of models is better in the southeastern parts of the study area. The area under the curve (AUC) analysis shows that all models struggle to forecast normal rainfall (with AUC <0.66) compared to above and below normal events. The ECCC, ECMWF, and JMA models exhibit better skills, while DWD and Météo-France models perform poorly in terms of both deterministic and probabilistic metrics. The multi model ensemble forecasts generally improve rainfall forecasts, reducing biases and improving correlations across all the months and regions compared to the individual models. The multivariate regression multi-model ensemble (MRMM) approach outperforms (having a perfect correlation value of 1 in region VI for example) the simple arithmetic mean (MMM) and the bias-removed multi-model ensemble (MBMM) in all metrics across regions and months. Seasonal rainfall forecasts show higher performance in low-rainfall months and regions, with notable challenges in wetter months and regions. The comparison between the EMI probabilistic forecast and the models indicated that the EMI forecasts performed better for above-normal rainfall categories, while the climate models were more reliable in predicting below-normal rainfall. Both the EMI and model tercile forecasts were limited in forecasting normal rainfall events. By identifying strengths and areas for improvement in current forecasting models, this research provides valuable insights for more accurate and region specific climate forecasts to allow for better preparedness and response to climate variability