National scale 3D modelling of soil organic carbon stocks with uncertainty propagation - An example from Scotland

- Use of a hybrid GAM-geostatistical 3D model
- Uncertainty propagation in 3D soil modelling
- Use of globally and freely available remote sensing data for DSM at national scale
- Improved agreement with validation set compared with existing DSM methods

The variation of soil properties down a profile is usually considered continuous. The aim of this study was to develop and test a methodology to model the continuous vertical and lateral distributions of SOC stocks in Scottish soils making explicit the modelling and spatial uncertainty of the results. A comparison with regression kriging and other depth function methods is provided to show that better performances can be achieved taking into account non-linear relationships between covariates and soil properties. The analysis was run for the whole of Scotland. The carbon stocks were calculated for each point, i.e. each horizon in each available profile. The stock value at each cell for each of the considered depth layers was defined using a hybrid GAM-geostatistical 3D model, combining: 1) the fitting of a GAM to estimate the trend of the variable, using a 3D smoother with related covariates; and 2) kriging or Gaussian simulations of GAM residuals as spatial component to account for local details. The use of GAM makes the approach flexible, because it is able to deal with both linear and non-linear relationships between soil properties and the considered covariates. The results confirmed that MODIS data are a useful source of information for DSM especially at national scale. When comparing the proposed approach with similar methods such as regression kriging, the results showed better agreement with the data in the validation set with a global R2 of 0.60. The median values obtained are comparable with the values reported from previous studies on stocks in Scotland using different methods. The uncertainty is large indicating a wide range of credible values for each pixel.