Associated Incidents
Ministers have been left red-faced after unveiling the first AI-generated map of Britain's peatlands --- only to have it comprehensively debunked by the people who live in the areas.
Local residents have reported that the government's state-of-the-art AI has confused prime peat with stone and walls, woods with degraded peatland, and even missed well-known examples of the genuine article.
The inaccuracies appear right across the country, from Devon to Shropshire and the Lake District. The technology even confused a wood --- first written about by William Shakespeare --- as prime peat bog.
It has led AI experts to warn that the government needs to proceed with far more caution in its headfirst approach to the new technology and adopt the old maxim of "rubbish in, rubbish out".
The peat map was released with a fanfare by the Department for Environment, Food and Rural Affairs (Defra) earlier this month, amid grandiose claims that the computer learning had, for the first time, been able to identify all areas of peatland across the country with a 95 per cent rate of accuracy.
It was hailed as a new way of using AI to provide a vital tool that could help policymakers combat peat erosion, reduce flood risk and prioritise funding to restore degraded areas that contribute to global warming.
It also contained a more sobering warning: 80 per cent of England's peatlands were in a dry and degraded condition and needed urgent attention to combat the threat of global warming.
The publicity around the new map was enough to prompt many farmers and conservationists to examine the peatland in their local areas. The map enabled viewers to zoom into as little as ten square metres.
But rather than being amazed they were left bemused by areas they knew well that were clearly misidentified.
Among those mystified was Cat Frampton, an organic beef and sheep farmer on Dartmoor, who has long been interested in preserving the peatland on her 100-acre farm.
"I did what you always do and zoomed in on my land," she said. "There is a steep field bang next to my house which is rocky soil and sand but the map showed it covered in peat. I then scrolled out and looked at other areas and it was the same.
"The map showed 87cm of degraded peat right on top of Haytor, which [has] massive lumps of rock sticking out of the ground, some of the most iconic bits of Dartmoor."
Frampton said she then turned on a filter to show "bare" peat --- and was able to spot lines of supposed deposits in exactly the same places where real maps showed dry stone walls, hedges and even trees.
"It was all comprehensively wrong," she said.
Frampton added that the problem was not just wasted effort --- but that policy decisions would be made off the back of inaccurate information.
"In somewhere like Dartmoor national park you need to know where to protect and where to provide money but you can't do that with bad information."
Frampton was not the only one to identify problems. Two hundred miles away Tim Ashton, who farms at Soulton Hall, near Wem in Shropshire, discovered that Soulton bluebell wood --- which inspired Shakespeare's* As You Like It* --- was mapped as prime peat, while a 40-acre field that he knew to be degraded peat was unlisted.
"On a public policy level this is just useless," he said. Farmers in Dorset, the Lake District and other parts of the country also reported problems, while the National Farmers Union said they had been getting a lot of feedback from their members on the problems.
David Exwood, the NFU deputy president, said: "Accurate baseline data is essential if this map is to guide effective policies for protecting peatlands while supporting sustainable food production.
"Without reliable data collection methods, farmers cannot have confidence in future land management policies."
Lily-belle Sweet, a researcher in the use machine learning for environmental science at the Helmholtz Centre for Environmental Research in Leipzig, said the problems arose because the AI did not have enough "real" data points about peat on the ground to "learn" accurately from.
She said this allowed the machine to misidentify the spatial data it was analysing, and that there were not sufficient real-life checks to pick up the mistakes and correct them.
"The moral of the story is that you will end up wasting time and money unless you work very directly on the ground by the people affected by your data so you know that it is accurate," she said.
Joseph Hillier, director of analysis at Natural England, said: "While this is the best tool we have to understand the location and features of peat across the country, it should be used alongside on-site verification when making decisions about our peatlands locally."