Resumen
Worldwide, machine learning (ML) is increasingly being used for developing flood early warning systems (FEWSs). However, previous studies have not focused on establishing a methodology for determining the most efficient ML technique. We assessed FEWSs with three river states, No-alert, Pre-alert and Alert for flooding, for lead times between 1 to 12 h using the most common ML techniques, such as multi-layer perceptron (MLP), logistic regression (LR), K-nearest neighbors (KNN), naive Bayes (NB), and random forest (RF). The Tomebamba catchment in the tropical Andes of Ecuador was selected as a case study. For all lead times, MLP models achieve the highest performance followed by LR, with f1-macro (log-loss) scores of 0.82 (0.09) and 0.46 (0.20) for the 1 h and 12 h cases, respectively. The ranking was highly variable for the remaining ML techniques. According to the g-mean, LR models correctly forecast and show more stability at all states, while the MLP models perform better in the Pre-alert and Alert states. The proposed methodology for selecting the optimal ML technique for a FEWS can be extrapolated to other case studies. Future efforts are recommended to enhance the input data representation and develop communication applications to boost the awareness of society of floods.