Explaining building exposure using urban morphology and AI

Exposure is a key aspect of the risk assessment framework. If we know the assets and people that are exposed to hazards, we can better estimate damage and losses in case of a disaster. Recently, building footprint datasets such as OpenBuildingMap have become available, that enable the quantification of exposure on a high resolution. While such datasets are nearing a global footprint coverage, the completeness of semantic information related to the buildings, such as occupancy type, is lower.

Machine Learning (ML) methods can be used to infer semantic information about buildings. Typically, a remote sensing approach is taken, where image-based ML techniques are used on satellite imagery. Such techniques require high-resolution imagery for good results, that are not widely openly available, and use a high amount of computing resources when covering large areas.

Rather than satellite imagery, we have used morphometrics to predict building information. Morphometrics are quantitative features that explain the structure of the built environment. They are calculated using solely building footprints and the street network; both are commonly available. The morphometrics exist on three building scales: the individual building (e.g. footprint size); the building plot (e.g. distance between neighboring buildings); and building block (e.g. building footprint coverage compared to the area of the block). There are also morphometrics related to the street network (e.g. network density). As each feature is a single value, the dimensionality of the feature space is much lower than for image-based methods, reducing the need for computing power. Using this method, we can predict building properties like its height, occupancy type and construction year.

This approach fills information gaps in existing building footprint datasets and can be integrated into high-resolution exposure modeling efforts like the Global Dynamic Exposure Model. It allows for the augmentation of heterogeneous building exposure in data-scarce regions where other methods, such as crowd-sourcing are not available. It can substantially reduce the otherwise high uncertainties in exposure modeling. Consequently, it supports decision-making at local, regional, and national levels, where authorities in civil protection must act despite incomplete or uncertain information.