Total soil carbon assessment: linking field, lab, and landscape through VNIR modelling

Context: Point based measurements provide only a limited overview of landscape variation in measured properties. Upscaling of measurements from point to landscape comes with challenges particularly considering error propagation. Objectives: We investigated the impact of using proximal derived measurements of soil total carbon taken at point locations on upscaling to landscape levels. Methods: 1087 soil samples across Florida, USA were collected, laboratory (LAB) analysed for total carbon (TC), and then measured using visible-/near-infrared (VNIR) spectroscopy. Proximal TC values were generated through chemometric modelling using random forest (RF) and partial least squares (PLS) regression. These three datasets were then upscaled to the State of Florida, USA using ordinary kriging and compared. Results: R2 (RPD) values for the PLS and RF chemometric models were 88% (2.96) and 91% (3.23), respectively. All 3 spatial models had an accuracy of 54% on an independent validation dataset, with greater than 70% accuracy if predicted values were considered within the interpolation variance range. When comparing spatial interpolations derived from the proximally measured samples, only 18% of the PLS versus 51% of the RF fell within a range of 0.05 logTC (g kg−1) of the LAB measured interpolations. Conclusions: Using proximal sampling and modelling provides comparable output to laboratory measured soil TC measurements at point level, but when upscaled to landscape level the selection of proximal modelling method will impact the spatial interpolations derived. The error propagation within sequential modelling must be considered particularly when one wishes to use sequential modelling to analyse change in environmental properties.