Accurate assessment of total carbon (C_t) content is important for fertility and nutrient management of soils, as well as for carbon sequestration studies.  The non-destructive analysis of soils by diffuse reflectance spectroscopy (DRS) is a potential supplement or alternative to the traditional time-consuming and costly combustion method of C_t analysis, especially in spatial or temporal studies where sample numbers are large.  We investigated the use of the visible to near-infrared (VNIR) and mid-infrared (MIR) spectra of soils coupled with multivariate chemometric analysis to determine their C_t content.  Our specific focus was on Hawaiian soils of agricultural importance.   

This study benefits from the remarkably diverse soils of Hawaii.  Our sample set included 216 soil samples from 145 pedons from the main Hawaiian Islands archived at the National Soil Survey Center in Lincoln, NE, along with 89 newly-collected samples from Kauai, Oahu, Molokai, and Maui.  In total, over 100 series from 10 of the 12 soil orders were represented.  The C_t values of these samples ranged from < 1% – 55%. 

We evaluated two methods in our effort to develop a prediction model for soils of unknown carbon content given only their spectral signature.  For each method, the calibration set was a random 70% subset of the full data set, and the validation set consisted of the remaining 30% of the data set.  Using Partial Least Squares Regression (PLSR), we found R² values of 0.91/0.94 (calibration/validation) for the VNIR spectra and 0.95/0.94 for the MIR spectra.  With a preliminary trial of random forests ensemble tree regression, we achieved slightly lower R² values for the VNIR spectra (~0.85) but comparable R² values for the MIR spectra (~0.96).  These results indicate that VNIR and MIR spectroscopy provide a viable alternative to combustion techniques for C_t analysis on Hawaiian agricultural soils.