Public Surveillance and the Future of Urban Pluvial Flood Modelling

Abstract

Motivation and objective Urban pluvial flooding is an issue of increasingly critical importance due to climate change and urbanization. However, the numerical models used for flood forecasting and risk mitigation suffer from a pronounced lack of monitoring data, which affects model accuracy. Monitoring data are necessary so the models, which contain undefined parameters, can be calibrated and validated against real flood events. In particular, it is important that the models are able to reproduce flood behavior in and around buildings, where the most damage is caused. However, conventional flow and water level sensors reach their limits in public spaces like streets due to irregular topography, moving obstacles, and the risk of vandalism. It has been suggested that surveillance cameras and social media could provide the necessary surface flooding data at a fraction of the cost of conventional sensors. The objective of this thesis is to explore how trend-like data can be extracted from surveillance footage and assimilated to boost the reliability of urban pluvial flood models. Methodology and outcome The first deliverable of this thesis addresses the general lack of data necessary for researching novel, video-based monitoring methods. For this, the floodX data sets were created by conducting realistic flooding experiments at a special flood training facility, where conventional sensors and surveillance cameras collected data in parallel. Besides supporting multiple research projects, including the methods developed and tested in this thesis, the rapid publishing of the data pushes the bar for Open Research in the field of urban water management. The second main deliverable of the thesis is a method (SOFI) to automatically obtain trend-like flooding data from surveillance cameras with the help of a convolutional neural network. The SOFI method is fully automatic, which will be necessary for deployment due to privacy concerns and data volume challenges. Despite the limited number of images used, it was possible to train a deep convolutional neural network to detect floodwater in a range of flooding situations. By performing an analysis every few seconds, trend-like water level data was obtained that had, on average, a correlation of 75% with the actual water level. While the SOFI method suffers if image quality is poor or if large obstructions block the view of the water, it has the advantage of being applicable to footage without the need for on-site surveys. In the third part of this thesis, practical value of trend-like data for improving a flood model’s predictive performance was assessed under a variety of data qualities and sensor network layouts in the floodX case study. The results indicate that error-free trend-like data can be nearly as good as - and sometimes better than - sensors when it comes to improving model performance. However, this result only holds for trend-like data with uncorrelated errors or no errors at all. Correlated errors in the trend-like data reduced the improvement achieved and in some cases, the use of erroneous data even worsened model performance. Based on the results, models calibrated with trend-like SOFI data still need to be cross-validated to ensure that any errors present in the data are not compromising predictive performance. Impact Although public surveillance is controversial from a social perspective, this thesis suggests that surveillance cameras could be used as a cost-effective data source to enhance the reliability of urban pluvial flood models, for the benefit of flood risk mitigation. For cities rushing to adapt to a changing climate with more intense rainfall, trend-like data obtained from surveillance footage, uncalibrated sensors, or other sources remain a promising alternative to costly conventional sensor data.