Mid-infrared attenuated total reflectance spectroscopy for soil carbon and particle size determination

In soil analysis mid-infrared spectroscopy (MIR) is not as widely used as visible and near infrared spectroscopy (VisNIR), mainly due to (1) the need for sample preparation and (2) strong absorption leading to spectral distortion and total absorption. In this study we used attenuated total reflectance (ATR) as a sample presentation technique to obtain MIR spectra of neat soil samples to overcome the two aforementioned drawbacks. A set of diverse soil samples (N=270) were scanned with a Fourier Transform IR spectrometer (4000 to 400cm−1) on its diamond ATR crystal. The objective was to investigate the usefulness of ATR spectra in determining clay, sand, organic and inorganic C using partial least squares regression. Results showed that inorganic C and clay can be predicted very well with ATR, with validation R2 around 0.94 and Ratio of Performance to Deviation (RPD) around 4.0. Sand can be predicted satisfactorily also, with R2=0.88 and RPD slightly lower than 3.0. Organic C is predicted with R2=0.77 and RPD>2.1. Compared to the results of VisNIR models, significantly higher accuracy was obtained for clay, sand, and inorganic C, and a slight improvement was obtained for organic C. In our opinion, MIR-ATR can be a promising and powerful tool for soil characterization. It combines the advantages of both VisNIR (minimum sample preparation and high analysis throughput) and diffuse reflectance MIR (better model performance). Finally, as evidenced in our “leave-whole-field-out” experiment, the models developed with ATR spectra performed better on dissimilar samples and might have a larger scope of application than VisNIR. •Obtained MIR-ATR absorption spectra of soil samples using an FT-IR spectrometer.•Associated absorption peaks in ATR spectra with soil chemical constituents.•Developed PLSR models of sand, clay, OC and IC using the ATR spectra.•ATR models performed superior than VisNIR models.•ATR models performed better on dissimilar samples than VisNIR models.