Digital soil mapping of peatland using airborne radiometric data and supervised machine learning – Implication for the assessment of carbon stock

Peatlands account for approx. 4.23 million km² of the land surface of Earth and between 5 % and 20 % of the global soil carbon stock, however much uncertainty exists. The release of carbon from modified peatlands is significant and affects the global carbon balance. The importance of conservation and rehabilitation of peatlands is clear. Global estimates currently use national scale mapping strategies that vary depending on available resources and national interest. The most up-to-date methods rely on satellite remote sensing data, which detect peat based on a multiband spectral signature, or reflected radar backscatter. However, satellite data may not be capable of detecting peat under landcover such as pasture or forest. Airborne geophysical surveys provide relevant subsurface information to update or redefine peatland extent maps at a national scale. Radiometric surveys, which measure the naturally occurring geologically sourced potassium, uranium, and thorium, offer the largest potential. Modelling of gamma ray attenuation shows that peat has a distinctive attenuation signature, due to its low bulk density, when considering all recorded radiometric data. This study exploits this signature by combining airborne radiometric data in a machine learning framework and training an artificial neural network to detect those data which have been acquired over previously mapped peatlands. A ∼95 % predictability is achieved. The trained neural network can be then used to predict the extent of all peatlands within a region, including forested and agriculturally modified peatlands, and an updated peatland map can be produced. This methodology has implications for global carbon stock assessment and rehabilitation projects where similar datasets exist or are planned, by updating the extent and boundary positions of current peatlands and uncovering previously unknown peatlands under forestry or grasslands.