Exposure Data

Why do we need high-resolution exposure?

When we calculate risk, we need to know where buildings are. In practice, it is rare that every building is individually included in a risk assessment framework. Instead, buildings are grouped together at a single location, such as a city centre, an administrative centroid, or a geocell. The problem is that hazards are variant in the spatial domain. Even earthquake shaking varies significantly across short distances, driven by local soil conditions that can amplify or dampen ground motion by a factor of two or more. When buildings are lumped together, the expected ground shaking they are assigned may not be representative at all for the actual ground shaking at the location of the building. The result is that some buildings are assumed to shake more (or less) than they would in reality and because of the nonlinear relationship between shaking and damage, these errors do not cancel out.

This bias is not random. Research consistently shows that coarser aggregation leads to underestimation of frequent, moderate losses and overestimation of rare, catastrophic ones. As is expected,l this problem grows with the size of the aggregation unit. Aggregation on state-level will experience larger biases than aggregation on the neighbourd level. This is why we need high-resolution exposure.

Why is our exposure model probabilistic?

For most cities, detailed structural surveys of individual buildings do not exist. Instead, we have the footprint of a building, its approximate height, and its occupancy type (residential, commercial, industrial, et cetera). From these attributes alone it is not possible to say with certainty what a building is made of or how it was designed. A mid-rise residential block could be reinforced concrete frame, confined masonry, or unreinforced masonry. A commercial building of the same height is far more likely to be reinforced concrete. The material is unknown, but it is not equally unknown for every building.

Our exposure model takes advantage of this. At the district level, the relative frequency of different structural types is known from aggregated surveys and census data. As an example, we know that 55% of residential buildings in a given area are masonry and 45% are reinforced concrete. Rather than assigning every building a single structural type (which would imply a false certainty) or ignoring building-level information entirely (which would discard heterogeneity in exposure), each building is instead assigned a full probability distribution over the structural types that are plausible given its occupancy and height. A residential building in that district might carry a 55/45 split between masonry and concrete; a commercial building of similar height, where concrete is far more prevalent, might carry a 15/85 split. The district frequencies provide the prior, and the building's own observable attributes filter and renormalise it.

Probabilistic risk frameworks are designed to propagate uncertainty through every stage of the calculation. From the hazard intensities, through the vulnerabilities, to the final loss estimate. A conventional exposure model that assigns each building a single deterministic structural type introduces a hidden assumption of certainty that is not justified by the data. It effectively treats classification errors as if they do not exist, which means the uncertainty in the final risk estimate is underreported. A probabilistic exposure model removes this inconsistency. Because each building carries a distribution over structural types rather than a single label, the uncertainty in the classification propagates naturally into the vulnerability step: buildings with ambiguous classifications contribute a range of possible fragility curves, weighted by their type probabilities, rather than a single curve.

Why do we need dynamic exposure?

District aggregate models are a huge effort to create and therefore are created only once every few years, meaning they go out of date. In a fast-changing world, the building stock changes by growing cities, redeveloped neighbourhoods and new constructions. The census data that aggregated models rely on, is collected infrequently. Especially in the developing world, a model built on a a decade-old census may reflect a city that no longer exists in quite that form, quietly introducing errors that are invisible until ground-truthed against real damage observations.

The Global Dynamic Exposure model build upon the aggregated models, which remain essential pieces in the puzzle. However, rather than relying on periodic large-scale surveys, it is built on top of OpenStreetMap, which is continuously edited. The code of the exposure model can be fully automated, meaning that it is easy to refresh the information for a newer date. Even though the model is based on aggregated models, that may have outdated information, we calibrate it to current population estimates on a country level. This means that in areas where we have seen more building development, there will be a higher population, even if this is not reflected in the census data.

A new release is produced every six months and each release is versioned, fully documented, and reproducible. Risk assessments can be tied to a specific snapshot of the world, compared across releases to track change over time.