Ensemble-Based Empirical Prediction of Ethiopian Monthly-to-Seasonal Monsoon Rainfall

Rainfall is the most important climate element affecting the livelihood and wellbeing of the majority of Ethiopians. The main Kiremt rainy season is from June-September (JJAS) and supports 85-95% of the country's food crop. Because all agricultural activities and resulting national crop production hinge on the amount and distribution of JJAS rainfall, accurate monthly and seasonal predictions of this rainfall are crucial for agricultural planning and disaster mitigation. A time-frequency based analysis was performed for concurrent teleconnections between monthly-to-seasonal Ethiopian rainfall and large-scale atmospheric circulation and global sea surface temperature (SST) patterns. This analysis linked Kiremt rainfall variation principally (66% explained variance) with annual time-scale atmospheric circulation patterns involving fluctuations in the major components of the monsoon system (e.g., monsoon trough, Somali low-level jet, tropical easterly jet). It also is shown that although variability on quasi-biennial (5%) and El Niņo-Southern Oscillation (2%) time-scales account for much less rainfall variance than the above annual mode, they significantly affect Ethiopian rainfall by preferentially modulating the major regional monsoon components through their season-long persistence. Because the effects of slowly evolving SST variations on Ethiopian rainfall must be realized through changes in local and regional circulations and oceanic patterns, a viable and successful monthly-to-seasonal Ethiopian rainfall prediction is best achieved by additionally identifying potential local and regional predictors and incorporating their aggregate effects by ensemble prediction methods. This study uses an ensemble-based multiple linear regression technique to assess the predictability of Ethiopian seasonal and monthly rainfall on local and national scales. The ensemble prediction approach seeks to capture potential predictive signals in regional circulations and global sea surface temperatures (SSTs) two to three months in advance of the monsoon season. A large pool of predictors (up to 71) were selected, from a larger set of orthogonalized variables, and used to initialize individual forward stepwise regression models. For retroactive validation (RV) approach, models are developed on 1970-1989 training data and forecasts are validated on mutually exclusive data for 1990-2002. For a leave one out cross validation (LOOCV) approach, predictor selection and model development processes are performed on training data from 1970-1999 that excludes a single year of rainfall data for forecast verification. For RV, the ensemble predictions reproduced well the observed all-Ethiopian JJAS rainfall variability two months in advance. The prediction outperforms climatology (persistence), with mean square error reduction, SS, of 46% (82%). The skill of the prediction remained high for the LOOCV approach, with the observed-predicted correlation (SSClim) being +0.77 (48%) for 30 verification years. For tercile predictions (below, near, above normal), the ranked probability skill score is 0.49, indicating improvement compared to climatological forecasts. At a local scale (Addis Ababa and Combolcha), the ensemble-average predictions provide skillful monthly rainfall forecasts with LOOCV correlation of +0.55. Compared to the previous generation of rainfall forecasts, the ensemble predictions developed in this study show substantial value to benefit society.