Natural England’s AI for Peat programme aims to restore carbon-rich peatlands with Microsoft Azure

The UK’s peatlands cover 12% of the country, locking in an estimated 3.2 billion tonnes of carbon. As well as being an important carbon sink, when in a natural, healthy condition, peatlands have a net cooling effect on climate and support biodiversity, including a unique range of flora and fauna. Peatlands also serve to reduce flood risk by slowing the flow of water from the uplands and providing floodplain storage in the lowlands.

What’s more, as in all ecosystems, peatland plants take CO₂ from the atmosphere. However, unlike most other ecosystems, this CO₂ isn’t fully released as peatland plants decompose. When peatlands are healthy, a substantial part of the carbon enters long-term storage in the peat.

Mapping is the first step to restoration

Only 22% of UK peatlands are in a near-natural condition. Degradation of the UK’s peatlands risks not only losing these important ecosystems and their environmental benefits, but also the release of a substantial amount of carbon into the atmosphere. It is estimated in England alone, this would amount to 584 million tonnes of carbon.

As a result, a new peat strategy is an important part of Defra’s 25-year environment plan. This includes funding for peatland restoration in England.

“England’s peatlands are degraded for any number of reasons, including overgrazing, afforestation, historic land management practices, drainage features, atmospheric pollution, and vegetation loss leading to erosion,” explains Karen Rogers, Lead Advisor in Natural England’s Cheshire and Lancashire Area Team. “However, we can’t start to restore them until we know their current condition and where any damage is located.”

A better way to plot peatland features

Mapping used to be a manual process, with teams walking the peatlands and plotting features. This was immensely time-consuming and precluded any kind of national mapping because of the sheer volume of work involved.

Even initiatives to plot peatland features from earth observation data such as Light Detection and Ranging (LIDAR) data were piecemeal, time-consuming, and often ineffective, requiring a lot of confirmatory groundwork. To map the entirety of England’s peatlands in this way would take approximately ten years of work. A more efficient method of mapping was required to enable a holistic approach to English peatland restoration.

In 2021, a proposal was put forward to use AI to map damage to English peatlands as part of the UK Civil Service Data Challenge, an initiative to promote better use of data across government.

Anne Williams, Project Manager for the AI for Peat programme at Natural England, explains the focus on innovation: “We would develop deep learning models to identify peatland surface features to better inform local restoration teams. By creating maps of the drainage features of grips and gullies, they can see where restoration efforts need to be focused.” Joe Hillier, Head of the Analytics Directorate for Natural England, adds, “This is life-saving work, so why wouldn’t we want to support that with the leading technology? We’re at the cusp of its potential—and that’s exemplified in projects like this.”

Grappling with huge datasets

The proposal won funding initially through the Civil Service Data Challenge and work began to develop the necessary models using aerial imagery and LiDAR data. AI4Peat’s data on peatland features and dams will be integrated into Natural England’s England Peat Map, due for publication in spring 2025. As well as the peatland drainage and erosion features, the wider map will also show the extent, depth, and condition of peat and peaty soils across England.

In standard satellite imagery, one pixel is equivalent to 10 metres squared. For Natural England’s mapping requirements, aerial imagery was required with a scale of 12.5 centimetres per pixel. The compute requirements to process imagery at that scale presented a big challenge.

Serendipitously, work was underway at Defra to develop a data analytics platform built on Microsoft Azure together with Databricks. It had been a long-held ambition to create a central analytics platform across the organisation, but it was during COVID-19 that the need to stand up analytics capability at pace became urgent.

Leveraging Defra’s Data & Analytics Science Hub

“Defra is a big user of geospatial data, so we needed a solution that could deal with that range and volume of data,” explains Paul Sinclair, Head of Data Exploitation at Defra.

Defra’s Data and Analytics Science Hub (DASH) team worked with Microsoft and Databricks to build a cloud platform on Microsoft Azure that could support the department’s data and analytics ambitions.

“We want to bring the whole analytics and science community together, supporting them to develop reproducible coding and analytical practices, scalable compute, innovation, and data science skills,” explains Olivia Newport, Head of DASH at Defra.

“We’ve worked closely with Microsoft and Databricks to stand up DASH,” explains Paul Sinclair. “Microsoft Azure was a very attractive choice for us because it hits our sweet spot in terms of cloud engineering and data engineering. It has been fantastic to lean into Microsoft’s and Databricks support to enable and accelerate this project.”

The DASH platform was launched in April 2022 and collaboration immediately commenced with Natural England’s AI4Peat team.

Collaborating on a Microsoft Azure cloud platform

The DASH platform was the perfect fit for the AI4Peat initiative. Close collaboration between the cross-governmental teams has been critical to the project’s success. The DASH platform provides new and necessary capabilities focused on scalability, agility, and performance, with Microsoft Azure tooling, machine learning, and AI capabilities creating a platform for collaboration.

“We just wouldn’t have been able to do this on a couple of networked power PCs,” explains Philip Shea, AI and Geospatial Lead at Defra and an original member of the Data Challenge team. “Our use of Microsoft Azure and Azure Databricks is enabling analysts to think in new ways. It opens up insights, allows us to scale our solution both spatially, temporally, and for other use cases across government.”

Joe Hillier, Head of the Analytics Directorate at Natural England, adds, “Natural England is a data-rich organisation and Microsoft Azure AI and Databricks AI capabilities are helping us get value from that data. Our ability to review that data and spot patterns is extremely limited as humans—so this type of project isn’t something we could have considered before this technology. The techniques and insights we’ve developed through this project have really moved us on strategically and in our capabilities.”

Boots on the ground

Once an area has been mapped by the AI4Peat AI models, the local teams can confirm and use the information to focus local peat restoration efforts where they are most needed and are likely to have the most impact.

“The deep learning models we have pioneered here are much more intelligent,” says Nick Tomline, Earth Observation Analyst at Natural England and an original member of the Data Challenge team. “They consider context, shape, and colour so we can map features much more accurately than what went before.”

This innovation is a huge boost to local teams working on the ground.

“We can load the output mapping layers onto a tablet and follow the features in the field,” explains Karen Rogers. “It makes everything so much easier. Instead of going out there to do the hard work of mapping features, we are asking a different question. We’re checking the details to confirm the mapping of the feature is correct and then asking, ‘What intervention should I be making here?’"

Perfecting the models

Even so, there are particular challenges with the datasets.

Martha Tabor, Data Scientist at Defra, explains, “In some areas, the vegetation looks very different which can make it harder to pick out the grips and gullies. Then there’s variation in the imagery itself; whether it’s taken at a particular time of day or from one season to another. We’re now thinking about how to ensure that the models can be accurately applied across the whole of England to take account of these variations.”

"We’ve achieved a lot in a short space of time,” agrees Joe Hillier, “but there’s a long way to go, of course.”

Exploiting the potential of AI for Peat

The potential to build on the results achieved so far is very exciting.

“We’re using semantic segmentation to identify grips and calculate their length, width, and depth. If we can now add some carbon metrics on top of that, we can start to get really useful digital datasets,” enthuses Anne Williams.

“I’d like to see the development of a digital twin through which we can model different types of interventions. This would enable us to see how the restoration of one area could affect another area, for example. We could share this type of consolidated map with our Area Teams, local councils, utility companies, and academic groups for further study.”

Enabling long-term monitoring and improvement

The estimated savings over a ten-year period is approximately £6 million, delivering outcomes that would have taken years to complete. On such a timescale, early insights would have been outdated before the project was completed.

“It might have been possible to do this manually, but this way we can update the work on a semi-regular basis,” emphasises Martha Tabor. “This will enable us to look at changes over time. A few years down the line, we’ll be able to plot where our restoration efforts are working.”

“I’d like to overlay these models with data about carbon loss and water loss,” Karen Rogers adds. “This could help us prioritise restoration work now and monitor change after restoration work is complete.”

All the data will be housed in a central repository to support peatland restoration. Over time, the project team also hopes to add data to indicate where restoration has taken place.

A new way of working to increase extensibility of models

Along the way, the team has reoriented to adopt best technology practices. One adaptation has been a shift to a more agile way of working and the creation of continuous integration and continuous delivery/deployment (CI/CD) pipelines to streamline and accelerate the development lifecycle.

“It’s been a bit of a 180˚ for us,” admits Anne Williams, “but working closely with Microsoft and Databricks has really helped us get the most out of these technologies. We have built our capabilities in collaborative tooling and created a more agile pipeline with links to all the code and artefact repositories where we store our models, datasets, and machine learning operations. This will enable us to easily reuse components and bring new components in to tweak and adapt our models.”

“We’ve already seen the benefit of this approach in AI4Peat. We have tested our grips and gullies model and retrained it to look for hags, another peatland feature,” Anne Williams continues. “We hope that, through our pipeline and repositories, we can share these assets and models for use on other projects across Defra.”

Olivia Newport confirms, “One of the really exciting opportunities of the AI4Peat project is how we extract our learnings and reusable components, products, bits of code and share those with other projects across Defra and other governmental departments.”

Sharing knowledge and collaboration

Collaboration has been critical to the project and platform’s success. Anne Williams states, “We’ve been able to push this work forward and break through barriers because of the cross-government working. There are pockets of deep learning expertise and AI skills across the Civil Service and we’ve been able to harness these different skillsets. It’s a true collaboration that has enabled us to get really creative and push things forward.”

Paul Sinclair agrees, “Success has also been driven by the team’s willingness to experiment. Accepting that no one gets things right first time all the time is really important. Being willing to take risks helps to push the boundaries of what’s possible.”

“We’ve had amazing collaboration from Microsoft and Databricks on the data science side and fantastic collaboration from Karen and the other Natural England teams on the ground,” confirms Nick Tomline.

Michelle Johnson, Data Scientist at Natural England, adds, “It’s been nice to know we have the support from Microsoft and Databricks, especially in scaling up and testing the capabilities at scale.”

A home for analytics expertise

The DASH platform is already home to other important projects, such as providing statistics for publication, green and blue space mapping projects, and the soundscapes project to analyse acoustic data in a range of natural habitats. It will be home to analysts working together to solve common problems and advanced AI projects.

Paul Sinclair explains, “The tools are available in Azure if we want to bring in additional libraries and languages. We have the ability to scale because we’re working with different partners on new projects. We now have the foundations; we’re providing a rich offer and that is expanding and extending all of the time.”

“Asking for help has also been important for our success,” adds Martha Tabor. “If you don’t ask, you don’t get. Asking for help and advice has led to a lot of the collaborations we’re benefiting from now.”

Innovation will continue

The multi-agency approach has delivered groundbreaking results, the benefits of which go far beyond the teams directly involved.

“I’m incredibly proud of the team. And the team should be very proud of what they have achieved. Work like this to mobilise natural capital will be hugely important going forward,” says Joe Hillier.

Olivia Newport agrees, “We’re all here because we want to make a difference. It’s really nice when our technical skills enable that. It’s so important and rewarding when those two things come together. Different parts of the organisation have come together to use emerging skills and technologies to solve intractable problems that it has not been possible to solve at scale before.”

“I think there will be lots of use cases like this which we will want to explore,” emphasises Paul Sinclair. “And we’re keen to work with Microsoft and Databricks to enable them. There’s a huge focus on AI and machine learning going forwards and our partnerships with Microsoft and Databricks will be absolutely critical to achieving those ambitions.”