Colombia, aided by aerial and terrestrial mapping data, detects buildings that may be vulnerable to disasters
Using artificial intelligence to analyze digital imagery collected through terrestrial and aerial mapping, the Capital District of Bogotá, Colombia, is working to identify homes and neighborhoods that need to be improved. The goal is to have families living in better, greener, safer homes before the next disaster strikes.
This effort, supported by the World Bank’s Global Program for Resilient Housing (GPRH), uses a Trimble MX7 vehicle-mounted mobile mapping system to capture street-level imagery across neighborhoods and combine it with aerial imagery collected by drones. (Personal data is not collected, and other measures, such as automatic face-blurring, are taken to protect privacy.) The datasets are then cleaned, organized and analyzed, with privacy protected and personal data not collected. Bogotá then uses machine learning algorithms to identify buildings and urban areas that could benefit from upgrading and an inspection by an engineer.
As Nadya Rangel, Secretary of Habitat of Bogota, points out: “This technology will help the city plan and implement future housing and urban upgrading investments in the most vulnerable areas.”
Making visible the invisible
The technology is making vulnerable areas more visible to decision-makers who determine public and private investment in housing and urban infrastructure. “Just as technology can help to detect cancer and provide treatment that we did not think was possible, governments can use technology to triage the homes and neighborhoods that could benefit from immediate upgrading interventions,” said Luis Triveño, senior urban development specialist and lead of GPRH.
Earthquake in Mexico puts focus on weak buildings
The idea of imaging neighborhoods to locate homes that are more vulnerable to earthquakes was sparked by the problem of soft-story buildings, said Sarah Elizabeth Antos, a data scientist and co-lead of GPRH.
A "soft story" is a structure in which one or more floors have windows, wide doors and/or large unobstructed commercial spaces or other openings where a shear wall would normally be required for stability against earthquakes. History shows that soft-story structures are prone to collapse during earthquakes and that seismic retrofitting is cost-effective. (Reference: FEMA (2012) FEMA p. 807: Seismic Evaluation and Retrofit of Multi-Unit Wood-Frame Buildings with Weak First Stories, Federal Emergency Management Agency (FEMA), Washington, D.C., United States.)
The risk of soft-story buildings was made evident in Mexico on Sept. 19, 2017, when a 7.1 magnitude earthquake caused more than 40 buildings to collapse, including the Enrique Rébsamen school, where 19 children and seven adults perished.
“It was very apparent from the street view imagery that this building was a soft-story. Looking at street view imagery and talking to engineers, it was clear it was a seismically vulnerable structure,” Antos said.
The World Bank GPRH team started to wonder: Could they use machine learning — computer algorithms that learn and adapt — to mimic an expert’s eye? Could they detect not just school buildings, but also types of homes that might be vulnerable to seismic activity so that engineers could be sent to conduct further assessments and governments can help retrofit them?
Bogotá neighborhood data informs investment
Now in its third phase of data collection, the Bogotá project is focusing on up to 12 vulnerable areas of the city comprising roughly 20 square kilometers.
The method generates a residential inventory of neighborhoods, or Geographic Information System (GIS) layers, that shows the government what might need intervention or upgrading, Antos said. “With this technology, governments can speed up the process of capturing data of the housing units in their districts so they can focus more on optimizing the use of scarce engineering resources and addressing the needs of vulnerable families.”
While mobile mapping and GIS technology are often deployed to create exposure databases or in response to earthquakes, floods or other natural disasters, they are also powerful tools for planning investments to mitigate risk before disasters strike.
“This type of work also creates and preserves the memory of how a place looked before a disaster or something bad happened,” Triveño said. Even when a natural disaster destroys an area, it helps to have the visual record. “That’s why we make available that imagery, so people can see it’s preserved, and it’s shared.”
Training the data to work smarter
In addition to donating equipment and software for the project, Trimble provided training on post-processing to the World Bank team and officials from the Bogotá Secretaría del Hábitat. Because of the pandemic, the training was done virtually and was led by a Trimble employee in Germany who spoke Spanish.
On the ground in Bogotá, the data collection team has included a driver, a passenger controlling the Trimble MX7, and someone in the office working with the software, which includes Trimble MX, Applanix POSPac Mobile Mapping Suite and Trimble Business Center. Privacy is an important consideration in the data collection process, so no personal data is collected, and faces are blurred. Once collected, the data is processed to improve its geo-position. Then, collected imagery from the MX7 is used to create panoramic images. The team then puts those images on Mapillary, an open-source, collaborative mapping tool, so they are viewable by the government and the OpenStreetMap community.
The World Bank applies its algorithms directly to MX7 imagery, and the machine learning process helps label and identify structures and characteristics that help determine if it is a poor or good building. In addition, the government, with some training provided by the World Bank, is providing aerial maps of the area for rooftop click-throughs that pull in the street-view images.
World Bank’s algorithms extract data and automatically categorize features of structures, such as:
- condition of the building, ranging from new to dilapidated;
- material used in construction;
- vintage or era of the building;
- architectural features, such as windows, doors and garages, which are a key indicator if a building is a soft-story structure; and,
- elevation to determine if a building has been raised, which would indicate an area prone to flooding.
The World Bank used a deep learning-based image segmentation technique, Mask R-CNN, to detect building instances from images and classify the detected instances regarding interested features, such as occupancy, material, and condition.
The instance segmentation technique is about categorizing and grouping pixels in an image into predefined classes. The Mask R-CNN method is essentially built upon the convolutional neural network technique, which has achieved significant successes in computer vision and image analysis in recent years.
The team also noted that the primary tool used is Detectron2, which is built upon and powered by the PyTorch deep-learning framework.
Ripple effects of Bogotá project
The World Bank is always testing new mapping and GIS technology, which continues to grow more accessible, with costs decreasing and ease-of-use increasing with better cameras and software.
“We have managed to bring it to a cost per unit that is very low, while keeping these algorithms open-source and trying to maximize the reach of these tools,” Triveño said. “I think in that sense, it’s very unique.”
“Looking at the big picture, the solution to vulnerable housing should be informed by technology; led by policies; and scaled up with private sector participation,” Triveño said.
This is Trimble-generated content and as such is not to be viewed as an endorsement by the World Bank Group.