Sufficiency of Features extracted from satellite remote sensing data for the estimation of local wind damage loss in urban areas

Kurzfassung

The damage done to buildings by meteorologic disturbances and their aftermaths, like wind-storms or flooding, is a major issue regarding entity insurance policies. While the functional model enabling the estimation of floodwater damage is well-known and accurate, an equivalent concerning storms and winter-gales is non-existent. To our knowledge the zoning system in operational mode currently is based mostly on the administrative boundaries of postal code districts. It is obvious that such a spatial segmentation does not represent the real hazard situation caused by wind-storms in any case. Especially in urban areas the small scale topography of buildings can generate amplification effects, i.e. in streets operating as wind channels, that largely exceed the windspeeds attained at laminar flow over a macroscale terrain. The need for the modeling of wind fields on a small scale is obvious. Modern optical remote sensing missions deliver images in very high resolutions which are able to provide suitable geospatial information on the layout and terrain of the land and man-made structures as well as the landuse of the observed earth’s surface. These can be exploited to generate the required wind parameters in a sufficient resolution to observe local effects that are most likely responsible for damage occurring during a strong storm event Further additional features concerning the object type of a particular structure can be extracted from remote sensing data, that play an imperative role in describing the impact of windspeeds of different magnitude on a building. As the damage function of wind events is a very complex one the demands on the Information inherent to possible features is very high. Therefore features produced from satellite remote sensing data are evaluated in this thesis regarding their capability of predicting monetary loss by building damage during a wind event. High-resolution wind and pressure fields are produced, along with building classes and vegetation density values to act as predictors for a non-parametric Regression model realized with boosted decision trees. The results of the machine-learning process and the succeeding validation show that a prediction of wind damage loss is not possible solely on the Basis of features extracted from remote sensing data, as they are not able to describe important building parameters that are essential for a successful estimation.