AI4Peat: Using computer vision to map England's damaged peatlands at national scale

The problem

Peatlands are the United Kingdom's largest natural carbon store. In good condition, they hold carbon in the ground, slow the flow of water off the uplands to reduce flooding, and support specialised wildlife. Decades of drainage, erosion and changing land use have left much of England's peat in poor condition, and repairing it is central to the climate and nature targets set out in the government's 2021 England Peat Action Plan.

Restoration begins with knowing the state of the land. That means locating the features that signal damage: grips, which are channels cut into the peat to drain it; gullies, natural channels that erosion has widened; and haggs, the steep exposed faces left where peat has worn away. It also means recording where repairs have already been made, which shows up in the small dams built to block drains and allow the ground to rewet.

This mapping was traditionally done on foot, with surveyors walking the land, or at a desk, with someone tracing features by hand from aerial photographs. Both approaches are slow and inconsistent, because different people record features in different ways, and covering all of England's uplands by hand was estimated to take around ten years. There was no single, consistent national picture of peatland condition, so decisions about where to restore, how to measure progress and whether policies were working rested on partial or out-of-date information.

The solution

AI4Peat is a project that uses artificial intelligence to map damage and repair across England's upland peat. It was led by Natural England with the Department for Environment, Food and Rural Affairs’s (DEFRA) Data Analytics and Science Hub (DASH), and won the Civil Service Data Challenge in 2021. It works by using computer vision, a branch of AI that lets software recognise and classify features in images.

The team trained two kinds of deep learning models, software that learns to spot patterns from many examples, on two sources of data: high-resolution aerial photographs at 12.5 centimetre resolution, where each pixel covers a 12.5cm square of ground, and LiDAR, which uses laser pulses to build a three-dimensional picture of the terrain. One model, using a method called semantic segmentation, learned to pick out grips, gullies and haggs and tell them apart from the surrounding land. The other, using object detection, learned to find the restoration dams.

The project built its training examples from mapping that peatland partnerships had already produced through fieldwork. Moors for the Future, the Yorkshire Peat Partnership and the National Trust supplied large amounts of labelled imagery, in effect showing the models what each feature looks like, with the England Peat Map team coordinating the effort and checking the material for quality and consistency.

Running the trained models across England's upland peat produced a national map of these surface features. The processing ran on DASH, a shared cloud platform built on Microsoft Azure with Databricks that handles heavy computation faster than an ordinary machine. Because the same analysis can be re-run when fresh imagery is captured, the map can be kept current as conditions change.

The impact

1. England gained a national map of upland peatland surface features.
Run across the country's upland peat, the models produced a national-scale record of where drainage channels, gullies and eroded edges lie and where restoration dams have been built. The outputs were published in May 2025 as part of the England Peat Map, which Natural England describes as the most detailed national coverage of England's peat to date. The tool has been operational since spring 2025.

2. The development cost was a fraction of the manual alternative.
The UK government's AI Knowledge Hub records the project's development at around £288,000, compared with an estimated £6 million to map the same area by hand over about 10 years.

3. After launch, the wider England Peat Map drew sustained criticism over accuracy.
Within weeks of the May 2025 launch, farmers and land managers reported inaccuracies, with peat shown in places it does not exist and missed where it does. Speaking to Envirotech Online, Dartmoor farmer Cat Frampton called the map “plainly wrong”: limestone pavements at Malham Cove were mapped as deep peat, reservoirs and rivers showed up as peat-rich soils, granite tors and quarries were marked as degraded bog peat, the vegetation layer placed bog plants on bare rock, and some known peatlands were entirely missing. These misidentifications fall within the extent, depth, and vegetation layers that the wider map builds from satellite and other predictor data. In contrast, the upland surface layers from AI4Peat showed a separate error, with evening shadows read as bare peat in some areas. Natural England responded that the models have an overall accuracy above 95 percent for the extent of peaty soils and 94 percent for vegetation and land cover, that some peat is missed. Some predicted where it may not occur, and that the map is a national-scale tool that needs local evidence for site-level decisions and should not be used to judge peat depth at a single site.

4. The map remained in use, and concerns continue to be voiced.
As of 2026, the England Peat Map remained published and openly available, integrated by the Forestry Commission and referenced in the government's Land Use Framework. Natural England attached technical disclaimers advising against site-specific use. Natural England's chair, Tony Juniper, has described the map as a basis for targeting restoration toward the most degraded peatland.

Lessons learned

• AI delivered national coverage that manual mapping could not. Walking the uplands or tracing features by hand could not produce a consistent national map within a usable timeframe, with the manual approach estimated at around ten years. The models applied the same criteria across the whole landscape at once. For a government that needs to monitor the environment over large areas, it shows how consistent national coverage can be reached where local surveys alone could not.
• Existing partner data was the foundation. Deep learning needs large numbers of labelled examples. The project used fieldwork that restoration partnerships had already done, so the models learned from features identified by people with direct knowledge of the land. For an agency weighing up AI for monitoring, checking what data partners already hold is a practical first step.
• Building it to repeat makes it a monitoring tool. Because the analysis runs on cloud infrastructure, it can be rerun on new imagery, allowing the map to show how conditions change over time. In environmental policy, tracking change is often a central requirement, and a repeatable pipeline supports that tracking.
• A high national accuracy figure does not guarantee local reliability. Natural England reported overall accuracy above 95 percent for peat extent, and 94 percent for vegetation, yet the errors were not spread evenly. They concentrated in terrain the models found hard, including rock outcrops, water, woodland and shadow, which is why critics found granite and reservoirs mapped as peat. The team had said before launch that differences in vegetation and in the timing and conditions of imagery make features hard to detect, and that the models needed testing across the country, especially in lowland peat. The response after launch showed the distance between a national-scale accuracy metric and what a user needs at a single site. Validating across varied terrain, stating plainly where the map can and cannot be relied on, and carrying those limits into the tools and policies that reuse it, are an important part of the work itself.
• The capability can extend beyond peat, with the same validation needs. The team noted that similar techniques could count animal species, track land-use change, or monitor coastal erosion, and that the platform and skills built for AI4Peat could support other projects across Defra and the wider government. Reusing the platform and skills can spread the value of the original investment, provided each new use is validated for its own terrain and purpose.

This case study was written with assistance from artificial intelligence.