Predicting Food‐Security Crises in the Horn of Africa Using Machine Learning

In this study, we present a machine‐learning model capable of predicting food insecurity in the Horn of Africa, which is one of the most vulnerable regions worldwide. The region has frequently been affected by severe droughts and food crises over the last several decades, which will likely increase in future. Therefore, exploring novel methods of increasing early warning capabilities is of vital importance to reducing food‐insecurity risk. We present a XGBoost machine‐learning model to predict food‐security crises up to 12 months in advance. We used >20 data sets and the FEWS IPC current‐situation estimates to train the machine‐learning model. Food‐security dynamics were captured effectively by the model up to 3 months in advance (R2 > 0.6). Specifically, we predicted 20% of crisis onsets in pastoral regions (n = 96) and 20%–50% of crisis onsets in agro‐pastoral regions (n = 22) with a 3‐month lead time. We also compared our 8‐month model predictions to the 8‐month food‐security outlooks produced by FEWS NET. Over a relatively short test period (2019–2022), results suggest the performance of our predictions is similar to FEWS NET for agro‐pastoral and pastoral regions. However, our model is clearly less skilled in predicting food security for crop‐farming regions than FEWS NET. With the well‐established FEWS NET outlooks as a basis, this study highlights the potential for integrating machine‐learning methods into operational systems like FEWS NET. Plain Language Summary In the face of increasing droughts and food crises, this study explored the use of machine learning to provide predictions of food crises in the Horn of Africa, up to 12 months in advance. We used an algorithm called “XGBoost,” which we fed with over 20 data sets of potential food security drivers. After training the model, we found that food security dynamics were accurately predicted up to 3 months in advance, especially in pastoral and agro‐pastoral regions. The model accurately predicted 20% of crisis onsets in pastoral areas and 20%–50% in agro‐pastoral regions with a 3‐month lead time. In agro‐pastoral and pastoral regions, our machine learning algorithm showed a similar performance to the established early warning system from FEWS NET. The machine‐learning model did not show good performance in crop‐farming areas. Nonetheless, this study underscores the potential of integrating machine‐learning methods into existing operational systems like FEWS NET. By doing so, it paves the way for