AI for Humanitarian Action with Microsoft

The need
A neighbourhood does not have access to information about specific vulnerabilities that they need to fix to reduce disaster risk. They also cannot decipher from the warnings they receive what the precise extent of the impending risk is. This makes informed pre-emptive action by individuals/local communities almost impossible.

Currently, there are two data sets that a family or neighbourhood can access to be informed about their vulnerability and about immediate threats – vulnerability maps, and hazard warnings.

The Vulnerability Atlas of India and its subsets provide information about earthquakes, cyclones, and flood risks from national to district levels. In addition, landslide information is also available at a similar resolution. These however are not of much use to families and neighbourhoods as their resolution is very poor and there are very significant variations within one category of vulnerability that they identify.

Hazard warnings primarily come from the meteorological department, some new private players, and government departments that monitor specific data such as river flows. The forecasts are however very vague in nature and do not reveal the exact impact that can be expected at a particular location, such as the level of flooding or the impact of storm winds.

As a result, the vulnerability information and disaster warning are often not even taken seriously, and loss of life and property takes place. This, though applicable for all disasters, is most visible in the case of annual floods across the country.

How we helped
Sunny Lives: AI for Humanitarian Action project will change the way neighbourhoods use disaster related information to make decisions and to act for avoiding loss of lives and assets. Information about their inherent vulnerabilities will be available with a very specific context of their region and surroundings. Disaster alerts and warnings will be in forms that will tell them exactly what is expected in their current location and when, and what immediate action they should take. This will be backed with community plans that would have been prepared in advance through a consultative process.

The proposed approach involves the processing of very large volumes of data, virtually impossible to be carried out with manual processes involved. With each house, school, and public building of the country being a pixel with specific and unique attributes, data on vulnerability and warnings need to be correlated and made available in an easy format. In particular, the warning data will need to be processed in near real-time.

Many of the attributes of location-specific vulnerability and anticipated disaster impact have to be derived from secondary information, such as the building material commonly found in a cluster/neighbourhood that tells us how it will behave in a storm or flood of a specific strength. These interpretations are based on principles that can be programmed into machine learning and can be extended over wide geographies, with improvements over successive cycles.

The project will be able to demonstrate with evidence, how better information can save lives and property, and change mindsets.