Predicting and communicating flood risk of transport infrastructure based on watershed characteristics

Abstract

This research aims to identify and communicate water-related vulnerabilities in transport infrastructure, specifically flood risk of road/rail-stream intersections, based on watershed characteristics. This was done using flooding in Värmland and Västra Götaland, Sweden in August 2014 as case studies on which risk models are built. Three different statistical modelling approaches were considered: a partial least square regression, a binomial logistic regression, and artificial neural networks. Using the results of the different modelling approaches together in an ensemble makes it possible to cross-validate their results. To help visualize this and provide a tool for communication with stakeholders (e.g., the Swedish Transport Administration - Trafikverket), a flood 'thermometer' indicating the level of flooding risk at a given point was developed. This tool improved stakeholder interaction and helped highlight the need for better data collection in order to increase the accuracy and generalizability of modelling approaches.

Keywords: Artificial neural network; Binomial regression; Flood prediction; PLS; Stakeholder interactions; Transport infrastructure.