Machine learning algorithm for flash flood prediction mapping in Wadi El-Laqeita and surroundings, Central Eastern Desert, Egypt.

In the work described here, flash flood prediction mapping for the Wadi El-Laqeita in the Central Eastern Desert of Egypt was established, using machine learning approaches involving two algorithms—extreme gradient boosting (XGBoost) and k-nearest neighbor (KNN). Flash flood driving factors, including elevation, slope, curvature, slope-aspect, lithological rock units, distance from streams, stream density, and topographic wetness index (TWI) were selected. Based on the machine learning models, the XGBoost and KNN algorithms were quite similar, in terms of variables importance, with distance from the stream network, slope angle, elevation, and stream density identified as the key driving factors, in order of importance. It is often difficult to allocate model parameter settings; therefore, hyper-parameter setting optimization was applied to improve model prediction performance. The models were trained using 70% flooding location and 70% non-flooding data, with the remaining 30% flooding and 30% non-flooding location data used for model and simulation result validation. The applied models exhibited accuracies of 90.2% and 80.7% for XGBoost and KNN, respectively, showing that the XGBoost algorithm performed better than KNN in this situation. Therefore, XGBoost was used in a powerful approach to flash flood prediction mapping, with the obtained predictions providing important guidance for decision-makers with respect to future study site development. [ABSTRACT FROM AUTHOR]