Cities must rethink their disaster risk management strategies
Cities all over the world are bracing themselves for the next natural disaster, and the trends are not in their favor. Over the last 30 years there have been over 11,000 major disasters, and the economic losses caused by natural catastrophes continue to increase, with storms, earthquakes, and floods making up nearly 90% of the losses. In 2017 alone, losses from disasters worldwide were over 325 billion USD, which is equivalent to an astounding 0.05% of the world’s GDP.
Unfortunately, the financial protection gap is also growing along with the increasing losses, where the current insurance and reinsurance markets are unable to keep up with the trends, leaving governments, businesses, and individuals to bear the financial burden of disasters. With higher frequency and severity of events and the looming uncertainty over the course of climate change effects, cities around the world are starting to rethink their disaster management practices. There is a growing movement to complement traditional emergency management with an enhanced resilience approach through preventative risk reducing actions and more efficient post-disaster recovery. However, to be successful, such an approach must rely on proven risk analytics and up-to-date built environment data, which is often the bottleneck in data-driven risk management.
Data-driven analytics are the key to unlocking urban disaster resilience
Today, we are living in the midst of a technology revolution, and disaster risk analytics have to be at the forefront. To answer critical questions about cities’ exposure to disaster risk at scale, we turn to a combination of AI-based technology and proven expert analytics. At the core of this initiative, we believe that municipalities, private entities, and citizens should all have access to a much richer catalog of data on the urban environment they live in and the risks that they face. So the idea is simple: build a platform with detailed information about every building and infrastructure component in the built environment, suitable to assess disaster risk. In case of earthquake risk, for example, the database should contain information about whether buildings are made of concrete, wood or steel, whether house’s walls or frames are providing lateral strength against seismic forces, and what seismic standard (if any) was used to design the thousands, or even millions, of buildings existing in modern cities.
Just a few years ago, such a project would be considered unfeasible, demanding resources that most countries and communities do not have. However, with recent developments in computer vision and the increased availability of satellite imagery, it is now possible to automate a large portion of this work. Microsoft recently demonstrated the potential of this technology by predicting and releasing 124 million building footprints for the United States, using computer vision and satellite imagery.
Given the success of building footprint prediction, we were curious whether other risk attributes could be predicted as well. Is it possible to predict the age, height, or structural type of a building? Can we predict whether buildings are being used for residential, commercial or industrial purposes? If these attributes are predicted at scale, they can be integrated into a pipeline for large-scale risk assessments at individual building scale. To show the potential of this technology, we trained a state-of-the-art deep learning model to simultaneously predict building footprints and risk attributes. The TernausV2 deep learning model is initially trained on the SpaceNet building footprint dataset, achieving state-of-the-art results for footprint detection. The model is then extended to predict 4 additional attributes that are key to risk assessment, including building height, age, structural type, and occupancy. We refer to this extended model as the RLabV2 model. Given a satellite image, this model predicts risk attributes for every pixel in the satellite image.
Once the initial predictions are made, we use the watershed algorithm to convert pixel-by-pixel results into building footprint polygons, which contain all relevant risk attributes. The polygons then enter into a risk assessment pipeline using state-of-the-art Monte Carlo simulation procedures for hazard and damage. The result is quantification of disaster risk to individual buildings at a city scale.
The built environment and earthquake risk in San Francisco
The technology is currently being deployed in the San Francisco Bay Area – an area of high earthquake risk which is located in the center of seven major faults. It is the ideal city to pilot the promising technology given the team’s local insight and expertise into earthquake risk in the area. The RLabV2 model was trained on building data from a Stanford University seismic risk study conducted by Anne Hulsey, Professor Jack Baker, and Professor Greg Deierlein. The model was then deployed to predict building footprints and risk-attributes in San Francisco. The city was mapped at a building-by-building level, including key risk characteristics, using nothing but satellite imagery.
The detailed database of the built environment then becomes an input into the risk assessment pipeline designed to predict detailed damage and financial losses expected as a result of all possible earthquake scenarios in the area. The study is currently being extended to all of the nine counties in the Bay Area and will become the first ever computer vision-aided risk assessment at scale.
What’s next? Computer vision-powered risk analytics for cities around the world
So what is next for this promising technology? The potential to conduct rapid high-resolution risk assessments using satellite imagery and state-of-the-art risk analytics opens up opportunities for cities around the world to understand their risks and begin taking proactive steps towards enhancing disaster resilience. This is particularly relevant for developing countries, which often operate under severe data constraints and have scarce financial resources for disaster risk mitigation, and unfortunately, whose development efforts are threatened by the economic burden of disasters. The computer vision-powered technology can help cities overcome the lack of data and prohibitive prices for data collection, leading to a better understanding of risk and enabling data-driven disaster risk management practices. We are currently taking the next step towards the ambitious goal of creating worldwide information system of the built environment and deploying the technology to perform high-resolution mapping of residential buildings in 29 cities across Europe and Central Asia.