Fast and operational building damage estimation tool for urban pluvial flooding

Estimating flood damage is crucial for both disaster risk reduction in the prevention phase and crisis management during flood events. While models for predicting damage from riverine floods are well-developed, tools for estimating damage from urban pluvial flooding are less advanced. This is an important gap, as heavy rainfall can lead to flooding in a wide range of locations, not just along rivers.Here, we present a new machine learning-based tool to quickly estimate building-level damage from urban pluvial flooding caused by heavy rainfall. Three key improvements are incorporated into this tool, compared to the traditional use of stage-damage models developed for riverine floods in dismissal of the flood pathway or the use of newer, overly complex models. First, it was trained on data specifically from known urban pluvial flood events, rather than relying on models developed for riverine floods, which can lead to more accurate damage estimates for this type of flooding. Second, the tool utilizes the XGBoost algorithm, a powerful machine learning technique capable of capturing complex non-linear relationships in the data. Third, the tool's modular design allows users to efficiently utilize available geographical information when making damage estimates by fixing the area of interest and reducing one step of the data preprocessing, towards providing results quickly enough for real-time forecasting applications. To address the common challenge of missing data, the tool uses smart random sampling techniques to impute required building-level features that are representative to known buildings affected by this flood pathway, reducing exposure bias.The performance of the new tool was evaluated in two case studies in Germany, involving approximately 2,400 and 17,500 buildings, respectively. The tool was able to provide damage estimates in 1.1 and 6.0 minutes on a standard laptop, representing a 2-3 fold improvement in speed compared to a baseline approach. Furthermore, to validate the tool, estimates were compared to a fully independent dataset. The new tool reduced the estimate error by a factor of 4.3 compared to employing a riverine flood damage model, demonstrating its improved accuracy for heavy rainfall flooding events, although generally showing overestimation.The new tool, named FlooDEsT – Flood Damage Estimation Tool, comprising the damage function and its application strategy, has shown improvements in computation time and performance at the first pilot studies. Its expansion to other flooding events and comparison with other damage datasets shall clarify its generalization power towards an improved estimation of building damage at urban pluvial floods.