Protecting important habitats means first looking at the landscape to see exactly where they are located. Aerial photography can help in this regard by giving us a bird’s-eye view. However, interpreting the thousands of images required to create large-scale, high-resolution ‘land cover’ maps – which distinguish between different types of land cover or habitat, such as forest and farmland – can be a painstakingly slow process. Now, though, a pioneering project that used AI to produce a land cover map of the Peak District National Park in central England demonstrates the potential of the technology for speeding up the process whilst offering unprecedented levels of detail. Information from the new map is already feeding into wildlife monitoring and landscape restoration plans in the park, whilst the same techniques could support conservation efforts elsewhere.
The project brought together data science expertise from The Alan Turing Institute, remote sensing expertise from Cranfield University and local environmental expertise from the Peak District National Park Authority, where senior data research analyst David Alexander led the effort. “Having that knowledge from all those backgrounds just really worked,” says Alexander, whose organisation was one of ten to be selected for a new Turing scheme aimed at tackling environmental and sustainability issues.
Machine mapping
Detailed mapping of large landscapes relies on aircraft equipped with specialist cameras, which offer higher resolution than satellite images and often better visibility, as photos can be taken below cloud cover. Traditionally, though, classifying land cover was painstaking work, involving manually drawing around and labelling features captured in thousands of photos. Land cover was last mapped this way for the entire Peak District in 1991, when Silsoe College (now part of Cranfield University) helped the Countryside Commission carry out a land cover survey of all the UK’s national parks. Researchers used what were then cutting-edge digital technologies to store and display the data, but it still took four years to complete the project.
So, when Alexander contacted Daniel Simms, a remote sensing researcher at Cranfield University, to talk about repeating the survey, he wanted to know if they could automate it to get it done quicker. At that time, Simms was experimenting with neural networks – machine learning models whose underlying structure is inspired by the organisation of the human brain – to help the United Nations map opium production in Afghanistan, and he suggested applying similar techniques to map UK habitats.
Via the Turing scheme (which combines the Turing Internship Network and Data Study Group initiatives), they also brought in intern Thijs van der Plas, at that time a doctoral student in neuroscience, to build the AI model for classifying land cover. Van der Plas already had deep expertise in neural networks from another context – analysing images of brain cells in mice – and, conceptually at least, there was some overlap. Both problems can be tackled using neural networks that group together and summarise the information in neighbouring pixels. This works whether the AI is looking at images of brain cells or landscapes. “We’re essentially trying to reduce the amount of pixels we need to process whilst retaining all the information in the image,” says Van der Plas.
Once they had built their AI model, the researchers taught it to recognise different types of land cover by feeding it ready-classified examples. It then processed the imagery for the entire Peak District in less than a day, producing a map that captures features as small as 12.5 x 12.5 cm for all 1,439 km2 of the national park – meaning even the smallest patches of habitat, down to individual trees, can be accurately detected. Alexander says this map is already guiding the Peak District’s landscape recovery plans, which are based around UK government funding for restoring degraded landscapes to more natural states that can benefit wildlife. It is also informing monitoring efforts for threatened species like water voles. “You expect to find water voles around water, but we think their habitats are a lot broader than that, so our land cover map has been used to indicate where we need to collect more data, which could include imagery from camera traps or DNA in soil samples,” Alexander says. “It could then also be used to model water vole populations and their movements at a landscape scale.”

Broad horizons
According to Simms, the model’s value is broader than just this one map. It offers the wider environmental community a consistent way of mapping habitats in exquisite detail, across large areas. “Someone could take this Peak District National Park model and then further develop it for their own use,” he says. As Alexander points out, the image datasets needed to do the mapping for other national parks already exist. The team’s code and methods are also freely available online.
Alexander and Simms continue to collaborate on AI projects and are drawing on what they have learned so far to tackle their next challenge: advancing from mapping land cover to tracking how land cover changes over time. This could be done using their existing approach, by comparing the land cover classification results from images taken at different times. However, any mistakes the AI makes could quickly stack up this way. So the researchers are planning a more accurate approach that detects change directly by comparing the images as it processes them.
As part of the Turing scheme, they presented this challenge – along with some of their imagery – to a Turing Data Study Group in May 2023. Data scientists at this expert ‘hackathon’ spent a week focusing on the problem, coming up with several new research directions for the team to pursue. “This process has been invaluable to an organisation like ours, which doesn’t have much resource for large-scale data exploration,” says Alexander.
Two of the team have also begun initiatives with other Turing researchers. Alexander is working with members of the former Living with Machines project – which focused on analysing data from the Industrial Revolution – to study the disappearance of historical landscape features like walls and hedgerows across the Peak District. Meanwhile, Van der Plas is adapting the AI techniques he originally learned in neuroscience to other environmental problems. His current project involves using remote sensing and volunteer-collected data to understand species diversity in butterflies and birds.
Find out more about the combined Data Study Group and Turing Internship Network scheme here.
Top image: Paul Daniels