University of Twente Student Theses


Using machine learning to predict maximum waterheight for flash floods

Kuik, Martijn Eduard (2021) Using machine learning to predict maximum waterheight for flash floods.

Link to full-text:
(only accessible for UT students and staff)
Abstract:Flash floods are one of the most dangerous types of natural disasters. High velocities and the often short time between rainfall and flood occurrence make them hard to predict, especially in an early warning context. For this, flash flood events can be modelled using a hydrodynamic flood model, but these are often complex and take a long time to develop. Moreover, simulations, depending on the study area, can take up to hours to complete. Therefore, machine learning (ML) is a good option, with significantly shorter runtimes and a less complex development process. Predicting flood extent with ML is currently being done in the form of (flash) flood susceptibility analyses. However, current methods are limited in the information they can provide, mainly due to limitations of the input data. The speed at which these flash floods occur and disappear can it difficult to collect detailed flood extent data. Therefore this research, in the Kyungu river basin in Malawi, assesses the feasibility of predicting maximum waterheight of flash floods with machine learning. This is done by developing a hydrodynamic flood model of the study area, after which the generated maximum waterheight maps of different return period events are used to train a Random Forest (RF), an Extreme Randomized Forest (EXRF) and an Extreme Gradient Boost algorithm (XGB). The ML model was based on a selection of thirteen fundamental for hydrodynamic modelling, predictors and was validated using 10 fold cross-validated R2, MAE and RMSE. The training consisted of 10 year, 50 year and 100 year return period events, after which a separate 20 year and an 80 year return period event were predicted. This research results show that the five most important features in predicting maximum waterheight were the DEM, Wetting front suction head (PSI), upstream cumulative (ups), saturated hydraulic conductivity (ksat) and mannings’ N (n). Using these predictors, the EXRF algorithm was the most accurate for the training dataset, with an R2 of 0.68. However, when predicting the 20 year return period and 80 year return period events, the accuracy decreased to an R2 0.58 and 0.91, respectively. Evaluation of the results suggests that higher accuracy is caused by the significant similarity between trained and predicted events. Therefore this study shows that it is feasible to predict maximum waterheight with ML but also concludes that it has not been proven to be feasible for use in an early warning and early action context yet.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Spatial Engineering MSc (60962)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page