|dc.description.abstract||The eastern African region continues to be extremely vulnerable to droughts. In Kenya alone, about 50 million people have been affected by droughts since 1983. Rainfall in the region is bi-modal, with the long rains in the boreal spring and the short rains in the boreal autumn. Seasonal forecasts of rainfall are essential to improve drought preparedness and mitigation for water managers and farmers. The aim of this study was to forecast monthly rainfall totals at a set of locations in the equatorial east African region, using an Artificial Neural Network (ANN) called a Long Short-Term Memory (LSTM). The input dataset consisted of seasonal forecasts (SEAS5) of precipitation and temperature from the European Centre for Medium-Range Weather Forecasts (ECMWF) in combination with several climate indices related to the El Niño Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD) and the Madden-Julian Oscillation (MJO). The performance was assessed by means of the WNDI (Weighted Non-Dimensional Index) and the anomaly correlation. Benchmarks for the evaluation included: climatology, precipitation forecasts by SEAS5 and a multi-variate linear model based on the same input data as the LSTM. In addition, the ability of the model to forecasts anomalous seasonal rainfall was assessed based on the 2x2 contingency table and associated scores (hit rate, false alarm rate and Clayton Skill Score). Sensitivity analysis was carried out to evaluate the relative importance of the features in the input dataset.
For all lead times, the LSTM outperformed both the linear model and SEAS5 in terms of both anomaly correlation and WNDI, although it is not clear whether or not this is statistically significant. Especially at longer lead times the LSTM shows improved performance relative to SEAS5, due to a good coupling with the climate indices. At a lead time of 4 months and in terms of WNDI and the anomaly correlation, differences in performance between the long rains and the short rains were small. The LSTM model showed slightly better performance with a WNDI of 0.75-0.85 (equal to an RMSE of 6-7% of the mean annual rainfall) and an anomaly correlation of 0.76-0.77, compared to the linear model. In the context of anomalous seasonal rainfall, the performance in the short rains was substantially better, on average, with hit rates between 40-50%, false alarm rates of 4% and Clayton Skill scores of roughly 0.7. The linear model performed better at forecasting below-normal rainfall in the short rains. The LSTM showed better performance in the long rains, especially when forecasting below-normal rainfall, with a hit rate of 42%, false alarm rate of 8% and a Clayton Skill Score of 0.55, on average. With regard to operational use, especially forecasts of anomalous rainfall are of interest, as they may be associated with the occurrence of floods and droughts. The model developed in this study is underconfident in forecasting these anomalies and are therefore not sophisticated enough for operational use. However, due to the observed low false alarm rates, it may still provide valuable information to farmers and water managers near the stations. Several suggestions are made for the improvement of the model developed in this study.||