摘要：Flood risk modelling aims to quantify the probability of flooding and theresulting consequences for exposed elements. The assessment of flooddamage is a core task that requires the description of complex flood damageprocesses including the influences of flooding intensity and vulnerabilitycharacteristics. Multi-variable modelling approaches are better suited forthis purpose than simple stage–damage functions. However, multi-variableflood vulnerability models require detailed input data and often haveproblems in predicting damage for regions other than those for which they havebeen developed. A transfer of vulnerability models usually results in adrop of model predictive performance. Here we investigate the questionsas to whether data from the open-data source OpenStreetMap is suitable to modelflood vulnerability of residential buildings and whether the underlyingstandardized data model is helpful for transferring models across regions. Wedevelop a new data set by calculating numerical spatial measures forresidential-building footprints and combining these variables with anempirical data set of observed flood damage. From this data set randomforest regression models are learned using regional subsets and are testedfor predicting flood damage in other regions. This regional split-samplevalidation approach reveals that the predictive performance of models basedon OpenStreetMap building geometry data is comparable to alternativemulti-variable models, which use comprehensive and detailed informationabout preparedness, socio-economic status and other aspects of residential-building vulnerability. The transfer of these models for application inother regions should include a test of model performance using independentlocal flood data. Including numerical spatial measures based onOpenStreetMap building footprints reduces model prediction errors (MAE – mean absolute error – by20 % and MSE – mean squared error – by 25 %) and increases the reliability of model predictionsby a factor of 1.4 in terms of the hit rate when compared to a model thatuses only water depth as a predictor. This applies also when the modelsare transferred to other regions which have not been used for modellearning. Further, our results show that using numerical spatial measuresderived from OpenStreetMap building footprints does not resolve allproblems of model transfer. Still, we conclude that these variables areuseful proxies for flood vulnerability modelling because these data areconsistent (i.e. input variables and underlying data model have the samedefinition, format, units, etc.) and openly accessible and thus make iteasier and more cost-effective to transfer vulnerability models to otherregions.
