Modelling soil carbon fractions with visible near-infrared (VNIR) and mid-infrared (MIR) spectroscopy

Accurate assessment of soil carbon fractions would provide valuable contributions towards monitoring in ecological observatories, assessment of disturbance impacts, global climate and land use change. The majority of chemometric modelling studies have focused on measuring only total soil carbon (C), with only a few evaluating individual soil C pools. Analysis of pools allows for a more detailed picture of ecosystem processes, specifically decomposition and accretion of C in soils. This study evaluated the potential of the visible near infrared (VNIR), mid infrared (MIR) and a combined VNIR–MIR spectral region to estimate and predict soil C fractions. Partial least squares regression (PLSR) and random forest (RF) ensemble tree regression models were used to estimate four different soil C fractions. The soil C fractions analysed included total — (TC), organic — (SOC), recalcitrant — (RC) and hydrolysable carbon (HC). The sample set contained 1014 soil samples collected across the state of Florida, USA. Laboratory analysis revealed the wide range of total and organic C values, from 1 to 523g·kg−1, with only about 10% of the samples containing inorganic C which was therefore omitted from the study. Both PLSR and RF modelling were shown to be effective in modelling all soil C fractions, with as much as 94–96% of the variation in the TC, SOC and RC pools, and 86% of HC being explained by the models. Although both PLSR and RF models were successful in modelling C fractions, RF models appear to target the physical properties linked to the property being analysed, and may therefore be the better modelling method to use when generalising to new areas. This study demonstrates that diffuse reflectance spectroscopy is an effective method for non-destructive analysis of soil C fractions, and through the use of RF modelling a spectral range between 2000 and 6000nm should suffice to model these soil C fractions.