In this guest post, Smart Data Research UK Fellowship recipient Dr Stef De Sabbata discusses her work integrating open data from the Geographic Data Service to produce a geospatial foundation model that reveals the population dynamics of places in England and Wales
Where should a local authority open a new childcare service? Which neighbourhoods would benefit most from a public health campaign, or have enough demand to sustain a new shop?
Answering questions like these requires combining a wide range of data about people and places, as well as modelling the complex relationships that connect them. Entities such as local authorities and small businesses seldom have the resources to engage with such modelling, nor to outsource it to specialised companies.
Through my Smart Data Research UK Fellowship, I set out to explore whether that heavy lifting could be done once, openly, and transparently, by creating a digital fingerprint of every neighbourhood in England and Wales. The result, the Output Area Foundation Model (OAFM), is an AI model, but not the kind that most people have in mind these days.
What are foundation models?
Over the past five years, the public imagination around artificial intelligence (AI) has been completely captured by the arrival of the current generation of chatbots and image generation tools. Suddenly, interactions that seemed to be in the realm of science fiction have become commonplace, bringing with them a range of discussions and viewpoints from the most utopian to the most dystopian.
At the same time, there is another side to the recent AI advances, which more rarely share the limelight, despite being as consequential.
These are AI models that use approaches similar to those deployed to create chatbots and image generation tools, but focus on capturing key patterns in specific fields of application. For instance, Google DeepMind recently released the third version of its WeatherNext weather model, while its AlphaFold series has completely revolutionised protein structure predictions. Reading about these tools might feel far less like a science fiction novel and more like a high school math class, but their quality and scalability, and relatively small resource requirements compared to AI chatbots, have led to significant advances in their sectors.
Among the different categories of models driving advances in AI, foundation models play a specific role. While most models are developed for a specific purpose, foundation models are designed to capture foundational patterns within a field that can then be used for a wide range of tasks.
In geography, we can now create models that capture how populations change over time and space, creating general digital fingerprints that can be used for analysis in urban planning, public health, economic forecasting and any other field that could benefit from a broad understanding of place.
Towards a population dynamics foundation model for the UK
The OAFM prototype, the main output of my Smart Data Research Fellowship, is a geospatial foundation model that generates digital fingerprints for all the Census Output Areas (and the larger Lower-layer Super Output Areas) in England and Wales. The fingerprints integrate five open datasets from the Geographic Data Service’s Data Catalogue, including:
- The 2021 UK Census
- Access to Healthy Assets and Hazards
- Broadband Speed
- Dwelling Ages and Prices
- Points of Interest
The OAFM models England and Wales as a geographical network, connecting areas that are close to one another or linked by commuting patterns. During training, the information flows through the geographical network, allowing the model to capture the nature of an area beyond the local demographic and amenities. The procedure is also designed to dynamically hide part of the data to force the model to learn how to fill in the gaps from what remains, and thus better understand the relationship between different components of the input data.
When mapped, the final fingerprints reveal coherent patterns. A first look shows a clear divide between urban and rural areas and how distinct London is from the rest of the country, but exploring the nuances of the model can be challenging as the digital fingerprints are highly multidimensional and collaborate with one another to capture complex meanings.
Deploying foundation models for the public good
Indeed, one could say that AI models are grown more than constructed, and while we design them to achieve certain targets, we don’t specify how to get there; they find their own way. Countless results have shown that this leads to better results for complex tasks than handcrafting every single aspect of a model, but it also raises a whole range of new issues.
That makes the creation of open-source models even more important for the public good, as an open-source approach allows other researchers (as well as users) to inspect every aspect of the model and how it was created. Moreover, a large component of the future work around geospatial foundation models should focus on explainability and interpretability approaches that allow us to better understand how a model works. These aspects are particularly important for local authorities and public institutions that can be held accountable for any decision made using foundation models, and thus might find it more challenging to rely on proprietary models.
More broadly, I believe the OAFM and any subsequent model developed based on it will have the greatest impact on small and medium enterprises (SMEs), as well as local authorities and public institutions. These entities are faced with questions that need complex modelling (from local planning to site selection), but seldom have the capacity. Open fingerprints encapsulate a wide range of information and relationships, compressed into a numerical form that can be readily used for modelling and creating insights.
My fellowship has now concluded, but the OAFM prototype is a first step towards a full-scale, open population dynamics foundation model for the UK, one that integrates smart data, captures change over time and supports transparent, explainable use. In my new position as AI Scientist at Geolytix, I hope that my work will make location planning available for a wide range of customers, allowing for more efficient modelling and site assessment. Whether you work in local government, business or research, I invite you to explore the prototype and get in touch.
Acknowledgements
This work was supported by the Economic and Social Research Council, Smart Data Research UK Fellowship, grant number UKRI4013. The data for this research have been provided by the Geographic Data Service (geods.ac.uk), a Smart Data Research UK Investment: ES/Z504464/1.
Disclaimer
The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy or position of the Geographic Data Service.
