Spatial classification of peatland status using remote sensing and random forest

12% of the disturbed peatlands in the world are known to contribute approximately 4% to global greenhouse gas emission, according to UNEP. Northern peatlands spread above latitude 45 N (e.g. in Canada, USA, Scandinavia and United Kingdon) are biomes where climate change may occur earlier and rapidly thus contributing to greenhouse gas emissions more than many other biomes around the world. Peatlands can be classified as either “intact” or “disturbed” to determine whether they could be CO2 sources or sinks yet, the classification process requires a remote sensing approach because of limited accessibility to physical locations and intensive nature of field mapping. The study presented here uses spectral reflectance from peatland surfaces, together with topographic and climatic properties of the environment to classify peatland areas across Scotland. Distinct spectral reflectance responses in visible red between 0.63 and 0.69 µm, near-infrared between 0.85 and 0.88 µm, and short-wave infrared between 1.6 and 2.2 µm of the optical electromagnetic spectrum, topography, climate and land surface temperature have been used to discriminate between peatlands. A random forest classifier was trained using a 70/15/15 train-validation split, to predict peatland status. The classifier achieved an overall accuracy (F1 Score) of 72%, with a class-level accuracy of 94% for Forested, 84% for Drained and Eroded, 67% for Modified, and 44% for Near-Natural Peatlands at 100m resolution.  Based on these results, a national Scottish peatland status map is modelled at 100-meter resolution, demonstrating the potential of using the model for large-scale peatland characterization. This work presents a remote-sensing-based classification framework to support peatland mapping and status monitoring.