It was a wet early autumn, with floods plaguing much of the country, when ideas started circulating concerning the creation of a model that could better identify those most vulnerable to these natural disasters.
This was the origin of what would become the Intelligent Flood Readiness Model, created by Snowflake in collaboration with Ordnance Survey (OS), that identified the 1.2 million buildings most at risk from flooding in England.
“We wanted to make a difference”, said Fawad Qureshi, Global Field CTO at Snowflake.
The AI-powered model combines OS’ thorough building datasets with a range of other government data and a textual analysis of Flood Risk Management Plans (FRMPs), to create a model which can easily be used to assess vulnerability at a national scale - with it specifically identifying buildings that are not only vulnerable to flooding but which are not covered by existing flood plans.
Those involved in developing the model at both organisations believe that their findings could have implications for how policymakers and civil servants prepare for flooding in the future.
__
This is not the first time Snowflake and OS have collaborated, with the two notably having produced ‘Slopey Roofs’, a demonstrator and concept which also used Met Office data to determine roof suitability for solar panel installation in different counties and regions of Britain, in the past.
However, this project went further in its analysis to combine OS data with data from an even greater number of other sources, producing an exceptionally comprehensive overview of the areas and buildings most affected by flooding and its after-effects.
It concluded that 68% of the 1.2 million undefended buildings at risk of flooding were located in deprived areas, which often lack the resources to help its residents recover effectively following a flood. The vast majority of vulnerable undefended buildings were at risk from surface flooding, rather than the more obvious coastal or river flooding, and areas around the east coast of Yorkshire and the Humber have the highest number of properties that fall into this category.
According to Qureshi, the biggest obstacle in developing the model was determining what other data to connect OS records with.
“We had more difficulties thinking of that than we did building the model”, he said.
In the end, the developers decided to cross-reference OS’ building datasets with England’s Indices of Deprivation, to get an insight into the relationship between physically vulnerable housing and social precarity.
Flood risk and flood defence data from the Environment Agency, as well as AI-driven text analysis of government documents relating to flood risk management plans, are then overlaid upon this.
In order to resolve the problem of combining several layers of data which did not all possess the same level of granularity by using the ‘H3’ system developed by Uber, which divides space into grids of hexagons at different levels of detail. This allows data at different granularities to be mapped onto the same grid.
The result is a model which mass processes and synthesises information, producing a semantic and natural language model which is easily navigable even for those without technical knowledge. One can ask it to locate general statements in flood risk management documents, to give flood risk analyses for specific locations, the reasons why so many buildings are not protected, and more.
OS' National Geographic Database is particularly useful here due to its trusted, authoritative nature, and its incredible detail - it comprises over 600 million geospatial features, updated 30,000 times a day. Following recent major data enhancements, it is the most detailed map of Britain ever created.
“We’re very keen on the democratisation of data,” Qureshi said.
“You don’t need to be a geek to use the model - if you understand domain context, you can work with it”.
According to Tim Chilton, Managing Geospatial Consultant at OS, this model conducts analysis with ‘speed’ and ‘agility’ at a time when many organisations are reducing their geospatial capability.
The model also shows its workings, something which Chilton argues allows for its users to have more trust in its technology - with trust, as he underlined, being fundamental in services of this kind.
However, he emphasised that the model in its final form ‘is not the ultimate authority’. Indeed different agencies have the chance to add additional data, such as local insights, to improve the model.
Nor does the model eliminate the need for a ‘human-in-the-loop’ in his eyes, with specialists being able to either use or challenge its insights, and it not offering all the information required for decision-making on its own.
“We still need the experts’, " he said.
“Fawad [Qureshi]’s offer to government ultimately was that this model is there for them to use and incorporate into their own governance and processes. So this for me is the beginning of the journey for flood risk modelling using AI rather than the conclusion.”
Qureshi agreed, adding: “What I say is ‘AI will provide answers, but humans are needed to ask the questions’”.
“This is not the final answer, but it can lead to the next question.”