Classifying Drought in Ethiopia Using Machine Learning

Abstract

This study applies machine learning to the rapidly growing societal problem of drought. Severe drought exists in Ethiopia with crop failures affecting about 90 million people. The Ethiopian famine of 1983–85 caused a loss of ∼400,000–1,000,000 lives. The present drought was triggered by low precipitation associated with the current El Niño and long-term warming, enhancing the potential for a catastrophe. In this study, the roles of temperature, precipitation and El Niño are examined to characterize both the current and previous droughts. Variable selection, using genetic algorithms with 10-fold cross-validation, was used to reduce a large number of potential predictors (27) to a manageable set (7). Variables present in ≥ 70% of the folds were retained to classify drought (no drought). Logistic regression and Primal Estimated sub-GrAdient SOlver for SVM (Pegasos) using both hinge and log cost functions, were used to classify drought. Logistic regression (Pegasos) produced correct classifications for 81.14% (83.44%) of the years tested. The variable weights suggest that El Niño plays an important role but, since the region has undergone a steady warming trend of ∼1.6°C since the 1950s, the larger weights associated with positive temperature anomalies are critical for correct classification.