What happened

Natural England’s AI4Peat project, funded by UK government department DEFRA, used machine learning models trained on aerial imagery, LiDAR, and field data to create a national peatland map.

Despite claims of over 95 percent accuracy, users quickly discovered glaring mistakes: quarries, bare rock, woodland, and even iconic granite tors were labeled as deep peat, while some genuine peat sites were missed entirely. 

Given the map is intended to inform peatland restoration funding, land management, and environmental policy, the errors are seen to have real-world consequences, with landowners fearing they could be denied funding or face regulatory decisions based on inaccurate data.

Why it happened

The primary cause was the reliance on AI models trained with limited or imperfect ground-truth data, combined with the inherent difficulty of distinguishing peatland features from above. 

The models sometimes confused visual cues - such as the colour and texture of stone walls or rocky outcrops - with those of peat bogs, especially in landscapes where peat is rare or modified by agriculture. 

While Natural England did conduct field checks, critics argue that not enough “ground truthing” or local verification was done before the map’s release. 

The statistical accuracy touted by the developers masks the severity of errors in rare or complex landscapes, where even a small percentage of mistakes can have outsized impacts.