Guest author Gianni Cristian Iannelli, CEO at Ticinum Aerospace llc, provides an overview of how in-situ data can clearly change the global perspective of large-scale exposure assessment.
The risk equation is composed of three main factors, i.e. exposure, vulnerability, and hazard, as widely known by those engaged in risk assessment. In this post I would like to write about exposure, with specific reference to large-scale mapping in urban areas. In this field, data is all! High quality data on buildings results in low uncertainty in determining their exposure. Specifically, an accurate risk model is obtained when the analyzed buildings are described by many risk-relevant features, such as occupancy type, shape regularity, number and size of windows, etc. Achieving such a complete description is out of reach in most cases, as all these diverse information items can hardly be found in consistent and accessible databases. Sub-optimal choices have to be considered on a broad range of approximation levels from brutal averaging to refined mapping schemes. Still, building-specific information, even on a very limited set of features, remains a desirable integration to an exposure model. One of these features is the number of floors, and I will focus on that.
Existing tools for assessing building height may not be reliable
The number of floors is important in determining, for example, important structural properties of a building when seismic hazards are concerned, or how much refuge room can occupants rely on at higher floors in case of non-submerging flooding events. Almost all of the available methods for extracting such parameters at a building level use an indirect approach. They first identify the building height, and then derive the number of floors using a simple rule-based approach (e.g. a floor is 2.5 meters high). Heights can be derived by simply measuring shadows in very high resolution optical images, or taking advantage of multiple views of the same three dimensional objects. Different methods use specific acquisition techniques such as LiDAR, or aerial oblique imagery. Many research papers also use ‘Synthetic Aperture Radar’ sources, which come with a typical side-view geometry. As one can imagine, the above mentioned techniques use expensive data (i.e. large-scale analysis is quickly unsustainable) and/or complex analysis techniques (i.e. results may not be as reliable as one would like). Moreover, a building height may be ambiguous in terms of corresponding number of floors, especially when different types of buildings are involved, or when floor height is not constant.
Other, simpler methods to try and fill the gaps are based on low-quality/outdated statistical data at regional or even national scale, but most of the time these turn out to be unreliable, especially in developing countries where registries are not constantly updated or in an electronic form.
Street level Pictures can enhance exposure assessment
Another source of data, which appears to be still largely under-exploited in exposure assessment, is represented by street-level pictures. This type of data offers a precious feature: it offers the chance to directly view the front of each building! Many companies are collecting these in-situ datasets globally, increasing the amount of captured data at a daily pace and offering them for a cheap price or even for free in some cases. Among these, there are: Google (i.e. Street-View), Bing (i.e. Street-Side), Baidu, Mapillary, OpenStreetCam, etc.
Successfully converting in-situ pictures to enhance large-scale exposure assessment
Ticinum Aerospacehas developed a framework capable of automatically retrieving in-situ pictures on a given Area of Interest (AoI), extracting risk-relevant features and then attaching ‘tags’ (in this case, the number of floors) to GIS buildings footprints layer. The core of the framework is deep-learning based, automated classification of street-level pictures; the classifying machine has been built by applying the popular Convolutional Neural Networks (CNNs) techniques in a smart and efficient manner. Once the CNN has been trained using a supervised approach, the framework is completely automatic and no human supervision is required.
An example of the results, applied in the San Francisco area (USA), is visible in the figure below. The building pictures have been retrieved from Google Street View. The GIS layers representing the streets and the buildings footprints have been collected from Open Street Map. The ‘number of floors’ attribute is retrieved from each image, and then accurately attached to the relevant polygon in the GIS layer reporting building footprints.
It must be noted that the retrieved values directly state the number of floors and, unlike typical data methods, no rule-based approximation is required. Furthermore, thanks to the front-view acquisitions, the framework can be modified to extract more risk-relevant features, such as occupancy type, average revenue level, materials, number of windows, etc.
In conclusion, this new type of in-situ data can clearly change the global perspective in large-scale exposure assessment. Risk models will achieve better accuracy levels and at a cheaper price!