Résumé
This study examines the relationship between GRACE terrestrial water storage (TWS) data and climate indices related to temperature and precipitation across four semi-arid African regions: the Mediterranean (MED), West Southern Africa (WSAF), East Southern Africa (ESAF), and Madagascar (MDG). In addition to mean temperature and total precipitation, ten climate indices from the Expert Team on Climate Change Detection and Indices (ETCCDI) were analyzed for their relevance to extreme rainfall and drought. Using a Long Short-Term Memory (LSTM) neural network, the study improves TWS prediction accuracy, with feature ablation assessing the importance of each climate index. The LSTM model showed strong predictive performance (R 2 = 0.73-0.91), identifying the Standardized Precipitation Evapotranspiration Index (SPEI) as the key driver of TWS variability. Specifically, 12-month time scale SPEI was dominant in MED, while 6-month time scale SPEI played a crucial role in WSAF, ESAF, and MDG. Regional trend analyses revealed persistent TWS declines in MED and MDG (2002-2022) due to intensified droughts. In WSAF and ESAF, trends reversed between 2002-2011 and 2012-2022, with early positive trends, particularly in northern WSAF, followed by intensified TWS losses.