Flash Floods Prediction Using Precipitable Water Vapor Derived From GPS Tropospheric Path Delays Over the Eastern Mediterranean

Abstract

A flash flood is a rapid and intense response of a drainage area to heavy rainfall events. In the arid and semiarid parts of the Eastern Mediterranean (EM) region, the spatiotemporal distribution of rainfall is the most important factor for flash flood generation. A possible precursor to heavy rainfall events is the rise in tropospheric water vapor amount, which can be remotely sensed using ground-based global navigation satellite system (GNSS) stations. Here, we use the precipitable water vapor (PWV) derived from nine GNSS ground-based stations in the arid part of the EM region in order to predict flash floods. Our approach includes using three types of machine learning (ML) models in a binary classification task, which predicts whether a flash flood will occur given 24 h of PWV data. We train our models with 107 unique flash flood events and vigorously test them using a nested cross-validation technique. The results indicate a good agreement between all three types of models and across various score metrics. In addition, the models are further improved by adding more features such as surface pressure measurements. Finally, a feature importance analysis shows that the most important features are the PWV values from 2 to 6 h prior to a flash flood. These promising results indicate that it is possible to augment the current flash flood warning systems with a near real-time GNSS ground-based data-driven approach as demonstrated in this work.