Flash flood predictions over the Eastern Mediterranean using artificial intelligence techniques with precipitable water vapor, pressure, and lightning data.

Abstract

Flash floods occur when heavy rain causes a fast and powerful flow of water in a drainage area. In the Eastern Mediterranean region, which contains arid and semi-arid areas, the location and timing of rainfall is the most significant factor in the formation of flash floods. Predicting when and where extreme weather events such as storms, heavy rainfall, and flooding are likely to happen is a key challenge in the effort to prevent natural disasters. Here, we present an improved version of a previous work by Ziskin and Reuveni, which investigated the use of precipitable water vapor (PWV) data from ground-based global navigation satellite system (GNSS) stations, along with surface pressure measurements to predict flash floods in an arid region of the eastern Mediterranean. The previous study involved training three machine learning models to perform a binary classification task, using multiple unique flash flood events and testing the models using a nested cross-validation technique. The results showed that the support vector machine (SVM) model had the highest mean area under the curve (AUC) and the lowest AUC variability compared to random forest (RF) and multi-layer perceptron (MLP) models.  When tested on an imbalanced dataset simulating a more realistic flash flood occurrence scenario, all models demonstrated a decrease in the false alarm rate (precision) with a high hit rate (recall) performance. In this study, we extend the previous work by integrating nearby lightning data as a new feature in our studied dataset. The inclusion of this feature is motivated by the observation that heavy rainfall, which can lead to flood events, is often accompanied before by an increase in lightning activity. The experimental results show that the adding a 24-hour vector of nearby lightning activity improves the precision score significantly.