ENHANCING EARLY WARNING SYSTEMS BY IMPROVING SUB SEASONAL TO SEASONAL RAINFALL FORECAST OVER ETHIOPIA

Abstract

ainfall during kiremt season is essential for Ethiopian agricultural operations. Improving sub seasonal to seasonal rainfall forecasting is important to understand the distribution, frequency, and intensity of rainfall in the country. This study was designed to enhance the current sub seasonal to seasonal rainfall forecast by characterizing rainfall, evaluating the European Centre for Medium-Range Weather Forecasts ERA5 performance, and applying multiple linear regressions (MLR) and the nonlinear autoregressive network with exogenous inputs (NARX) model. Daily gridded rainfall data recorded over 184 station points was analyzed to identify rainfall onset, cessation, and length of growing periods. On the other hand, monthly gridded rainfall and four predictors (SST3.4, ERA5 total precipitation, wind at 850 and 200 mb level) covering the period of 1990 to 2020 were used to improve sub-seasonal to seasonal rainfall forecasting and for model evaluation. The model's performance was evaluated using RMSE, MAE, PBIAS, NSE, and R2. The result of this study indicates that the mean kiremt rainfall onset begins southwestern early and progressively spreads across the country's central, western, and northeast. This study generally finds that earlier onset dates for regions northeast, Central, and east, northwest has a later average onset in the current study, while southwest is similar but ends earlier. South highlands show a later start but an earlier end in the current study, whereas the southern has an earlier start with a shorter duration. Northeast, northwest, southwest, east, south highlands and south regions, consistently shows a later rainfall cessation. The standardized Kiremt rainfall anomaly results show that the majority of the country was under dry conditions during El Ni̱o years and wet during La Ni̱a years. The gridded data and the ERA5 model's output from the ECMWF ERA5 reanalysis demonstrated a strong correlation, but the model's underestimating rainfall value at both seasonal and sub-seasonal levels compared to the actual across Ethiopia. The R2 results are 0.87 and 0.86 for the MLR and NARX models, respectively, which indicate that the MLR model is slightly better suited to the observed data. This study found that both MLR and NARX models demonstrate outstanding skill for rainfall prediction for both seasonal and sub-seasonal levels of Kiremt seasons. The NARX model shows strong results in capturing the rainfall patterns in Ethiopia, particularly during wet seasons, and can provide valuable insights for understanding and predicting Kiremt rainfall in the country. The MLR and NARX models�۪ performance varies across different stations and periods. These results highlight the importance of considering different models and their strengths and weaknesses when estimating kiremt month levels in specific locations. The results of this study will improve early warning systems over Ethiopia in the fields of agriculture, hydrology, and other related fields.