Item

Abstract

Socioeconomic livelihoods in the Horn of Africa (HA) are highly dependent on seasonal rainfall, which occurs during two main seasons: October-November-December (OND) and March-April-May (MAM). During the two last decades the HA region has been affected by severe and prolonged droughts, leading to acute food insecurity, shortage of drinking water, and increasing risk of disease. Sub-seasonal drought prediction over the HA, from two weeks to two months, is therefore crucial for decision making and early warnings across several sectors. The sub-seasonal prediction of high and low precipitation extremes (PEs) by dynamical forecast systems is challenging for both rainy seasons, but there may be potential for extending the current prediction timescale based on remote drivers. To investigate the sub-seasonal predictability of PEs during the OND season we build a Long Short-Term Memory (LSTM) Neural Network predicting biweekly precipitation tercile categories over the HA region. The LSTM is trained on observational and reanalysis data during the period 1981—2020 and provides predictions with lead times of one week to one month. The results show that floods can be more skillfully predicted than droughts for all lead times. Moreover, we use explainable AI methods to explore the contribution of remote drivers to the predictions and potential sub-seasonal forecast opportunities for PEs. Preliminary results show that the sea surface temperature over the tropical Pacific is important for the LSTM prediction, but further investigation is needed to determine more factors affecting the prediction skill for PEs over the HA region.