Extreme Rainfall Events Classification Using Machine Learning for Kikuletwa River Floods

Abstract

Advancements in Machine Learning techniques, availability of more data-sets, and 1 increased computing power have enabled a significant growth in a number research areas. Predicting, 2 detecting and classifying complex events in earth systems which by nature are difficult to model 3 is one of such areas. In this work, we investigate the application of different machine learning 4 techniques for detecting and classifying extreme rainfall events in a sub-catchment within Pangani 5 River Basin, found in Northern Tanzania. Identification and classification of extreme rainfall event 6is a preliminary crucial task towards success in predicting rainfall-induced river floods. To identify 7 a rain condition in the selected sub-catchment, we use data from five weather stations which have 8 been labeled for the whole sub-catchment. In order to assess which Machine Learning technique 9 suits better for rainfall classification, we apply five different algorithms in a historical dataset for the 10 period of 1979 to 2014. We evaluate the performance of the models in terms of precision and recall, 11 reporting Random Forest and XGBoost as the ones with best overall performance. However, since the 12 class distribution is imbalanced, the generic Multi-layer Perceptron performs best when identifying 13 the heavy rainfall events, which are eventually the main cause of rainfall-induced river floods in the 14 Pangani River Basin.