Microsoft : Using AI to protect vulnerable communities from the effects of climate change

AI to the rescue

"We didn't have another instance of AI being used to tag roof types to forecast damage due to hurricanes. In addition, there was no readily available training data," says Tina Sederholm, a senior program manager in the AI for Good Research Lab at Microsoft, who led the project with data scientists.

"From a technical standpoint too, it was difficult because there is no urban planning in areas that we were targeting, and the population was so dense that it was difficult to first differentiate individual houses and categorize them accurately based on their roof type. But we built a machine learning model to counter these problems," explains Md Nasir, a data scientist and researcher in the AI for Good Research Lab.

To create the much-needed training data, Gramener, with its expertise in geospatial solutions, stepped in to deliver a scalable solution. Its data scientists accessed high resolution satellite imagery and manually tagged more than 50,000 houses to classify their roofs under seven categories depending on the material used to construct them.

"We wanted to identify the building footprint and distinguish between two houses distinctly. But informal settlements do not often have well defined boundaries and they are generally the worst impacted in any disaster," says Sumedh Ghatage, a data scientist from Gramener, who worked on building the AI model. "Secondly, as the geographical location changes, the types of roofs change as well. But we wanted to identify all kinds of roofs, to ensure the final model could be deployed in any region."

This formed the basis of the training data Nasir required. After trying a few different techniques, his final model could identify roofs with an accuracy of nearly 90%. But that was just the beginning.

"Apart from roofs, we considered nearly a dozen critical parameters that determine the overall impact cyclones would have on a house," says Kaustubh Jagtap from Gramener, who led the data consulting bits for the project. "For example, if a house is closer to a water body, it would be more likely to be impacted due to a cyclone-induced flood. Or if the area around the house is covered by concrete, the water won't percolate into the soil below and odds of water logging and flooding would be higher."

The team at Gramener then added other layers to the model. The alignment of all the different layers including road networks, proximity to water bodies, elevation profiles, vegetation, among others was a tedious task. Gramener created an Azure machine learning pipeline, which automatically captures the data and produces risk score profiles for every house.

It took about four months for the Sunny Lives model to become a reality and it was piloted during cyclones that hit southern Indian states of Tamil Nadu and Kerala in 2020. But it was during Cyclone Yaas in May this year that it was deployed at scale.

As soon as the path of Cyclone Yaas was predicted, the team at Gramener procured high resolution satellite imagery of densely populated areas that'd be impacted and ran the Sunny Lives AI model. In a few hours, they were able to create a risk score for every house in the area.

Gramener also assisted in sampling techniques and validated the accuracy of the model with actual ground truth information.

"Earlier, we used to deploy volunteers who manually conducted surveys. Now, all we need to do is procure high-resolution satellite imagery, run the model to determine an area's vulnerability and get the risk score results within a day. This kind of capacity was unthinkable earlier," says Garg.

Once the houses were identified, SEEDS along with its on-ground partners fanned out into the communities and distributed advisories to nearly 1,000 families in local languages like Telugu and Odia, which is spoken by the residents. Each advisory had detailed instructions on how they could secure their homes and where they would need to relocate to before the cyclone made landfall.

The model has opened a world of possibilities. SEEDS believes it can be deployed in many countries in Southeast Asia that share similar dwellings and communities that face the extreme levels of storm risk.

It can also be used to cope other weather challenges. For instance, SEEDS is looking at using the model to identify homes in densely populated urban areas that might be susceptible to heatwaves as temperatures hit new records every summer.

"During a heatwave, roofing becomes the most important parameter because maximum amount of the heat gained in the house happens through the roof. Houses with tin sheets often have poor ventilation and are the most vulnerable at this time," explains Garg.

There are other projects being piloted too. For instance, they are looking if AI could be used to identify vulnerable houses in the Himalayan state of Uttarakhand, which is prone to earthquakes.

"We brought our disaster expertise to the table, but Microsoft's data science made it possible for us to develop the model from scratch," says Ranganathan.

"The Sunny Lives AI model that the SEEDS and Gramener teams have created is a leading-edge humanitarian solution that is already saving lives and helping to preserve the livelihoods of people most at risk of natural disasters," says Kate Behncken, vice president and lead of Microsoft Philanthropies. "The ingenuity and collaboration between these teams is impressive, and I am encouraged by the promise that this solution holds to help better protect people for other severe weather scenarios, such as heat waves. This is exactly the kind of impact we're looking to support and drive with NGO partners via the AI for Humanitarian Action program."

Inspired by the results, SEEDS has started building its own technical capabilities after receiving the AI for Humanitarian Action grant from Microsoft.

"At the end of first year, we also started getting consultants to maintain and improve the accuracy of the model. Microsoft has given us access to the source code, so we may reach a stage soon where we will be able to run the model ourselves," adds Ranganathan.