A machine learning approach to flood severity classification and alerting

Abstract

With increased big data and computing power, machine learning is predominantly used for classification to object detection and forecasting of phenomena and relationships. Forecasting, mapping and impact assessment of flood events is one such area where machine learning is gaining momentum. While machine learning has been widely used for forecasting of flood extent and depth using rainfall/runoff datasets, impact assessment based on flood severity distribution using machine learning is still a long way from maturity. In this study, we used several machine learning classifiers such as Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), Support Vector Machine (SVM) and Multinomial Logit (ML) to classify flood severity into four classes: Information, Advisory, Watch and Warning based on the training datasets obtained from the Model of Models. The Model of Models is an ensemble model which integrates flood forecasting models to determine flood severity globally at sub-watershed level based on spatial extent and duration of flooding, risk scores associated with historic flooding events. The severity classes are used to disseminate alerts to stakeholders globally. The initial results reveal that the GB followed by DT and RF classifier performed better for classifying severity based on the performance assessment metrics. While this study has implemented a first version machine learner, future advancements will focus on deploying adaptive learners to increase the forecasting ability of the machine learner with new datasets generated daily.