Machine learning algorithms are excellent at answering “yes” or “no” questions. For example, they can scan huge datasets and correctly tell us: Does this credit card transaction look fraudulent? Is there a cat in this photo?
But it’s not only the simple questions – they can also tackle nuanced and complex questions.
Today, machine learning algorithms can detect over 100 types of cancerous tumors more reliably than a trained human eye. Given this impressive accuracy, we started to wonder: what could machine learning tell us about where people live? In cities that are expanding at breathtaking rates and are at risk from natural disasters, could it warn us that a family’s wall might collapse during an earthquake or rooftop blow away during a hurricane?
By 2030, 80% of the world’s population will be living in urban areas, following the dream of better jobs, education, and health care.
Too often, however, that dream risks remaining an urban daydream, due to natural disasters such as hurricanes, earthquakes, and floods, as well as climate change. Those of us working to help these families find a better future must focus more on ways to support efforts to protect their lives – and their livelihoods.
In the 40 years since the launch of Habitat I, governments and municipalities throughout emerging and developing countries have been proving that their cities can be not only inclusive and secure, but also resilient and sustainable. However, unless they increase their speed and scale, they are unlikely to achieve the goals of the “New Urban Agenda” and its Regional Plans, launched at Habitat III in 2016.
From our perspective helping governments in Latin America and the Caribbean, and ahead of the World Urban Forum taking place in Kuala Lumpur, Malaysia in February, let us share three key ingredients necessary to achieve that goal:
- UN Habitat
- City Resilience Program
- Global Goals
- natural disasters
- disaster risk management
- World Urban Forum
- New Urban Agenda
- Habitat III
- Sustainable Communities
- Social Development
- Urban Development
- Climate Change
- Latin America & Caribbean
- East Asia and Pacific
Can we rely only on satellite? How accurate are these results?
It is standard practice in classification studies (particularly academic ones) to assess accuracy from behind a computer. Analysts traditionally pick a random selection of points and visually inspect the classified output with the raw imagery. However, these maps are meant to be left in the hands of local governments, and not published in academic journals.
So, it’s important to learn how well the resulting maps reflect the reality on the ground.
Having used the algorithm to classify land cover in 10 secondary cities in Central America, we were determined to learn if the buildings identified by the algorithm were in fact ‘industrial’ or ‘residential’. So the team packed their bags for San Isidro, Costa Rica and Santa Ana, El Salvador.
Upon arrival, each city was divided up into 100x100 meter blocks. Focusing primarily on the built-up environment, roughly 50 of those blocks were picked for validation. The image below shows the city of San Isidro with a 2km buffer circling around its central business district. The black boxes represent the validation sites the team visited.
|Land Cover validation: A sample of 100m blocks that were picked to visit in San Isidro, Costa Rica. At each site, the semi-automated land cover classification map was compared to what the team observed on the ground using laptops and the Waypoint mobile app (available for Android and iOS).|
The buzz around satellite imagery over the past few years has grown increasingly loud. Google Earth, drones, and microsatellites have grabbed headlines and slashed price tags. Urban planners are increasingly turning to remotely sensed data to better understand their city.
But just because we now have access to a wealth of high resolution images of a city does not mean we suddenly have insight into how that city functions.
The question remains:
In an effort a few years ago to map slums, the World Bank adopted an algorithm to create land cover classification layers in large African cities using very high resolution imagery (50cm). Building on the results and lessons learned, the team saw an opportunity in applying these methods to secondary cities in Latin America & the Caribbean (LAC), where data availability challenges were deep and urbanization pressures large. Several Latin American countries including Argentina, Bolivia, Costa Rica, El Salvador, Guatemala, Honduras, Nicaragua, and Panama were faced with questions about the internal structure of secondary cities and had no data on hand to answer such questions.
A limited budget and a tight timeline pushed the team to assess the possibility of using lower resolution images compared to those that had been used for large African cities. Hence, the team embarked in the project to better understand the spatial layout of secondary cities by purchasing 1.5 meter SPOT6/7 imagery and using a semi-automated classification approach to determine what types of land cover could be successfully detected.
Originally developed by Graesser et al 2012 this approach trains (open source) algorithm to leverage both the spectral and texture elements of an image to identify such things as industrial parks, tightly packed small rooftops, vegetation, bare soil etc.
What do the maps look like? The figure below shows the results of a classification in Chinandega, Nicaragua. On the left hand side is the raw imagery and the resulting land cover map (i.e. classified layer) on the right. The land highlighted by purple shows the commercial and industrial buildings, while neighborhoods composed of smaller, possibly lower quality houses are shown in red, and neighborhoods with slightly larger more organized houses have been colored yellow. Lastly, vegetation is shown as green; bare soil, beige; and roads, gray.
Want to explore our maps? Download our data here. Click here for an interactive land cover map of La Ceiba.
Between 2005 and 2014, due to natural disasters, the region had a nominal cumulative loss of around US$5.8 billion, and witnessed more than 3,410 deaths and hundreds of thousands of displaced people. More recently, in October 2011, Tropical Depression 12-E hit the coasts of El Salvador and Guatemala with damages amounting to nearly US$1 billion.
In two recent studies, we evaluated the causal impacts of hurricane windstorms on poverty and income as well as economic activity measured using night lights at the regional and country level. In both cases, we applied a fully probabilistic windstorm model developed in-house, and calibrated and adjusted it for Central America. The first study (on poverty) used yearly information at the household level (for income and poverty measures) as well as the national level (GDP per capita). Due to the limited comparable household data between the countries, we decided to follow up with the second study (on economic activity) using granular data at the highest spatial resolution available (i.e., 1 km2) to understand more deeply the (monthly) impact over time.
Our results are striking: