Bayesian downscaling of building exposure models with remote sensing and ancillary information

Abstract

<p>Residential building exposure models for risk and loss estimations related to natural hazards are usually defined in terms of specific schemas describing mutually exclusive, collectively exhaustive (MECE) classes of buildings. These models are derived from: (1) the analysis of census data or (2) by means of individual observations in the field. In the first case, expert elicitation has been conventionally used to classify the building inventory into particular schemas, usually aggregated over geographical administrative units whose size area and shape are country-specific. In the second case, especially for large urban areas, performing a visual inspection of every building in order to assign a class according to the specific schema used is a highly time- and resource intensive task, often simply unfeasible.</p> <p>Remote sensing data based on the analysis of satellite imagery has proved successful in integrating large-scale information on the built environment and as such can provide valuable vulnerability-related information, although often lacking the level of spatial and thematic resolution requested by multi-hazard applications. Volunteered Geo Information (VGI) data can also prove useful in this context, although in most cases only geometric attributes (shape of the building footprint) and some occupancy information are recorded thus leaving out most of the building attributes controlling the vulnerability of the structures to the different hazards. An additional drawback of VGI is the incompleteness of the information, which is based on the unstructured efforts of voluntary mappers.</p> <p>Former efforts have been proposing a top-down/bottom-up approach moving from regional scale to neighbourhood and per-building scale, based on the analysis and integration of different data sources at increasing spatial resolutions and thematic detail. Following the same principle, this work focuses on the downscaling of already existing building exposure models based on census data making use of a probabilistic approach based on Bayesian updating. Different aggregation models can be taken into account to increase the spatial resolution of the building exposure model, also including variable-resolution models based on geostatistical approaches. Land-use masks are first generated after a supervised classification of Sentinel-2 images, in order to better relate the built- up area to meaningful geographical entities. Two independent statistical models are then created based on prior input information. Maximum likelihood estimations are obtained for each model. Two types of auxiliary data have been employed in order to constrain the downscaling via a specific likelihood term in the Bayesian updating: 1) building footprints area from the open-source-volunteered geo-information OpenStreetMaps  and 2) built-up height and density estimators based on remote sensing developed by the DLR (the German Aerospace Agency).</p> <p>This approach, developed within the scope of the RIESGOS, was tested in Valparaiso and Viña del Mar (Chile) where the residential building exposure model proposed by the GEM-SARA project has been downscaled. The performance of the different auxiliary data were separately tested and compared. An independent building survey has also been carried out by experts from CIGIDEN (Chile) using a Rapid Remote Visual Screening Survey and used for preliminary validation of the approach.</p>

Towards a Sensitivity Analysis in Seismic Risk with Probabilistic Building Exposure Models: An Application in Valparaíso, Chile Using Ancillary Open-Source Data and Parametric Ground Motions

“…The study area included the communes of Valparaíso and Viña del Mar (Chile), which have 295,113 and 326,759 inhabitants, respectively (2017) [51], forming the second-largest urban area of the country after Santiago de Chile. The area also houses the main port of the country, and therefore its physical security is vital for the national economy.…”

“…First, the urban area mapped from the land use land cover analysis (Section 3.1, Figure A1) was the external boundary of the grid. Then, the population reported in the official 2017 census data [51] was used to obtain the number of buildings of every typology based on their associated nighttime residents. This census data provides this type of information for various aggregation boundaries, ranging from small urban blocks up to large rural areas with few structures and inhabitants.…”

“…In this framework, emerging technologies have been proven to be useful to classify large-scale building stocks. For instance, we can mention the use of machine learning techniques over local datasets [22], feature extraction from remote sensing (e.g., [23,24]), façade image reconnaissance analyses (e.g., [25,26]), integrating multi-source information from remote sensing and surveys [27], and downscaling existing models based on remote sensing products being used as ancillary data (e.g., [28,29]). Moreover, methods to design focused surveys for data collection were also proposed (e.g., [30]).…”