Spatial variability and relative influence of seasonal rainfall drivers in Ethiopia

Rainfall variability is shaped by large-scale climate drivers, regional weather patterns and local land characteristics. While significant progress has been made towards understanding processes that contribute to rainfall variability, the degree to which these processes influence regions is contingent upon the specific climate region and its temporal dynamics. Existing predictive models tend to be generalized and do not fully account for the spatial variability present over small scales posing challenges for risk mitigation and tailored intervention. In this study, we use a random forest regression algorithm to identify hidden patterns and relationships between remote and local predictors influencing seasonal rainfall. Our selected case study area is Ethiopia, located in East Africa, where rainfed agriculture is predominant and seasonal rainfall is used for livelihood sustenance. We developed regime and season specific models using globally available rainfall and climate data. The model showed strong performance across all regions with a coefficient of determination ranging between 0.71 and 0.93. Amongst remote predictors, Indian Ocean Sea surface temperature (SST) indices were dominant during autumn, Pacific Ocean SST, and zonal wind during spring and Pacific SST, zonal wind, and Atlantic SST during summer season. A Tree Ensemble Interaction Quantification score showed that a pair of features (remote-local) together contribute more information to the model's prediction than each feature alone. For statistically significant interactions, we used Shapely Additive Explanations to evaluate the marginal contribution and nature of interaction across observations. The results suggest that climate models need to be tailored to regions to account for interactions and to improve reliability.