Impact-based prediction of building damage from surface water floods using machine learning.

In Switzerland, surface water floods (SWF) account for approximately 23% of the financial losses to property caused by floods. Improving the understanding of these events is therefore essential to enhance prevention and risk mitigation efforts. However, SWF impacts are challenging to forecast, as they result from the interaction of multiple processes and are strongly influenced by local conditions, building exposure, and vulnerability.

We develop a data-driven model to predict potential damages, trained on damage data provided by the Swiss Mobiliar Insurance Company and the Building Insurance of the Canton of Zurich (GVZ). The objective is to predict the probability of damage to buildings caused by SWFs using gridded hourly precipitation data and morphological properties.

We compare several approaches, including a simple threshold-based method, logistic regression, random forests, and deep learning models such as Convolutional Neural Networks (CNNs) and Transformers. The relevance of spatio-temporal patterns in precipitation fields is assessed using 1-D, 2-D, and 3-D CNNs. Variants of Transformer architectures are also evaluated.