Mapping Peat Depth Using Remote Sensing and Machine Learning to Improve Peat Smouldering Vulnerability Prediction

Abstract

Peat is an accumulation of soil formed from partially decomposed organic matter. Peat can burn, especially in hot, dry weather which is happening more often due to climate change; smouldering releases stored carbon to the atmosphere. Peat that has higher organic bulk density and lower moisture content is more vulnerable to fire: it will burn more severely (more deeply) if ignited. Shallower peat is less able to retain moisture during droughts and is therefore likely more vulnerable to fire; however, mapping peat depths at high spatial resolution is expensive or requires extensive fieldwork. This project uses remote sensing in combination with machine learning to estimate peat depth across a peatland and rock barren landscape. A Random Forest model was used to map peat depths across the landscape at a 1 m spatial resolution using LiDAR data and orthophotography. The resulting map was able to predict peat depths (R2 = 0.73, MAE = 28 cm) and showed that the peat depths which are especially vulnerable to high severity fire are distributed in numerous small patches across the landscape. This project also examined peat bulk density and found that the Von Post scale for peat decomposition can be used as a field method for estimating bulk density (R2 = 0.71). In addition, in this landscape, peat bulk densities at the same depth (within the top 45 cm) are higher in shallower peat because in shallower peat, more decomposed peat was found closer to the surface, and because peat with high mineral content was found close to the bedrock or mineral soil. The findings of this project will be valuable for wildfire managers to determine which areas on the landscape are most vulnerable to fire, allowing them to mobilize resources more rapidly for wildfire suppression.