Unravel High Resolution SOC Patterns in Agricultural Peatlands by Integrating Proximal and Remote Sensing into Machine Learning

Peatlands represent one of the largest terrestrial carbon pools worldwide. They are highly important for climate as they might act as huge CO2 sinks or sources depending on land use and management. This emphasises the significance of mapping the soil organic carbon stocks (SOCstocks) in peatlands. However, mapping SOCstocks is a tough task for the scientific community. Proximal sensing (e.g. soil apparent electrical conductivity) from electromagnetic induction (EMI) and high resolution remote sensing data (e.g. LiDAR and RapidEye satellite images) can support digital mapping of SOCstocks. The soil apparent electrical conductivity (ECa) has been a useful tool to estimate not only peat thickness but, SOCstocks. A few studies to date have attempted to use the digital soil mapping framework to map ECa data from EMI. This study evaluates the application of random forest algorithm through a digital soil mapping framework to map ECa and then use it as an environmental variable to increase models’ predictive power of SOCstocks in peatlands. To test this hypothesis, we applied three different scenarios defined as: spectral indices calculated from (1) averaged multitemporal image (e.g. single image); (2) non-averaged multitemporal images (e.g. six images); and (3) non-averaged multitemporal images of RapidEye satellite collection with terrain derivatives from LiDAR sensor as environmental variables to map ECa. Last but not least, we evaluated the use of the best fitted model for ECa from those three scenarios using its predicted map as an environmental variable to map SOCstocks in peatlands. The scenario 3 outperformed the other two ones (RMSE = 6.41 mS m-1, and R2adj = 0.95). Whilst, scenario 2 (RMSE = 7.11 mS m-1, and R2adj = 0.94) presented better metrics than scenario 1 (RMSE = 9.73 mS m-1, and R2adj = 0.90). In this sense, it was possible to evaluate the potentiality of adding the ECa information to increase modelling predictive power of SOCstocks, which presented better accuracy then not adding that information. Therefore, high resolution remote and proximal sensing data can assist to characterise and extrapolate the ECa and SOCstocks information through machine learning and digital soil mapping in peatlands to unknown areas overcoming the limitations of geostatistical methods. This study can provide new insights regarding the use of ECa data from on-go EMI for mapping peatlands and its properties