Disaster Planning with AI | A SEEDS & Microsoft Case Study

Predicting Disasters and Saving Lives
With Seeds and Microsoft

Challenge

SEEDS wanted to make disaster warnings and other risk-related information up to date and obvious down to a neighborhood

Approach

SEEDS partnered with Microsoft and Gramener to create an advanced machine learning-based model that could help them plan and respond more effectively to disasters

Outcome

SEEDS can recommend preventive actions to reduce the impact of any incoming disaster for the at-risk population

Partnership - seeds & gramener join hands

SEEDS is a not-for-profit organization that helps make communities resilient through comprehensive interventions in the areas of disaster readiness, response, and rehabilitation.

Gramener has partnered with SEEDS and Microsoft to develop the Sunny Lives AI model under Microsoft’s global program ‘AI for Humanitarian Action.

This predictive model aids in planning better risk reduction strategies against worsening climate emergencies and disasters.

Technology used for this solution

Creating the Sunny Lives model needed work from scratch due to some challenges:

The population in those areas was dense enough to make accurate differentiation between houses difficult. As geographies change, so do the roof types, and identifying these different roof types was of prime importance for the model’s concept.
There was no readily available training data to tag roof types for forecasting damage intensity.

Gramener, in partnership with Microsoft, built a Machine Learning model to counter these issues. To create the training data, Gramener accessed high-resolution satellite imagery and manually tagged over 50,000 houses. The roofs were classified under 7 categories based on the material used for their construction.

The final model could identify roofs with up to 90% accuracy. On top of this, Gramener added other geographic layers or parameters to the model.

The following parameters were considered – Waterbodies – River, Lake, Pond, and Sea, Distances from Road Network, Topographic Wetness Index (TWI), Elevation and Slope Vegetation (NDVI) – Sparse, Moderate, and Dense, Impervious surface, Landslide Risk, Building Footprints.

Developing the model took about 4 months. It was piloted in 2020 during cyclone Nivar and Burevi that hit the southern Indian states of Tamil Nadu and Kerala, respectively. The solution was deployed at scale in Puri during Cyclone Yaas in May 2021.

The solution required using a technology that could provide information that can be easily comprehended by the disaster response teams and the affected communities.

Considering that many disaster-prone areas are off the technology grid with low mobile penetration and network connectivity, it was necessary to break down the output from complex technologies like machine learning and artificial intelligence into simple and widely understood information. For example, SEEDS created specific advisories for dwelling types, suggesting the course of action for the individuals to reduce the risks.

A mock-up of the advisory shared by SEEDS with at-risk families. The solution utilizes high-resolution geospatial imagery and open datasets. It also uses geospatial analytics tools such as Kepler and Microsoft Azure to host and represent the results.

THE PROBLEM

When disasters like floods and heatwaves occur, warnings and other risk-related information are often vague and not up to date. Much of the current risk information work at a macro-level, covering hundreds of square meters of area and being too hard to understand by at-risk populations.

THE APPROACH

SEEDS partnered with Microsoft and Gramener to create an advanced machine learning-based model that could help them plan and respond more effectively to disasters.

This model uses historical data and satellite imagery to predict hyper-local risk information for early intervention. The Sunny Lives model builds on the knowledge that SEEDS has garnered over the years. The underpinning logic of the model is that the roofing material of a house can act as a proxy for its socio-economic condition. Hence, the capacity of a family living in a makeshift metal sheet home to respond and recover would be considerably less than that of one living in a large house constructed of concrete.

When two such dwellings exist in the same area, the impact of destruction caused by a disaster is significantly different for both. The solution maps this information on satellite imagery and other geographic parameters – forming the backbone of this AI solution.