Grinding and spectra replication often improves mid‐DRIFTS predictions of soil properties

Abstract

There is an increased interest in using diffuse reflectance infrared Fourier transform spectroscopy in the mid‐infrared region (mid‐DRIFTS) for high‐throughput prediction of soil properties, but basic methodological factors toward this end have yet to be thoroughly vetted. This study aimed to determine how the combined effects of soil grinding (sieved to <2.0 mm and finely ground to <0.5 mm) and sample replication (single or multiple soil subsamples, using one‐to‐four replicates) affect the spectral quality and predictive performance (accuracy) of automated plate‐based mid‐DRIFTS for soil analysis. We evaluated chemometric prediction performance of soil physical, chemical, and biological variables (clay, sand, pH, total organic C, and permanganate‐oxidizable C [POXC]) in 397 soils from the U.S. Midwest. Sieved soils (<2.0 mm) increased the overall spectral variability compared to finely ground soils (<0.5 mm) and led to a distinct wavenumber importance allocation in support vector machine models. These spectral changes degraded prediction performance of <2.0 mm samples when compared to <0.5 mm samples. The number of spectral replicates had a smaller effect on spectral properties, but impacted prediction accuracies of soil properties. In general, prediction outcomes improved with four spectral replicates either within a single soil subsample or across different soil subsamples. Our data collectively suggest that soil particle‐size reduction to <0.5 mm and collecting multiple spectra improve mid‐DRIFTS predictions. Recommendations to optimize high‐throughput mid‐DRIFTS should consider the tradeoffs between prediction accuracy and the effort needed to prepare soil samples and acquire spectra.

Fine grinding is needed to maintain the high accuracy of mid‐infrared diffuse reflectance spectroscopy for soil property estimation

“…The superior model accuracies with fine grinding could be attributed to several factors. First, the homogenization process and finer soil particles can lead to a more uniform distribution of material in the sample holder, enabling the capture of more representative reflectance even with a small beam spot size of the MIR instrument (Deiss et al., 2019; Guillou et al., 2015). This can be important for reducing errors in calibration transfer efforts to use models calibrated on one instrument setup to another.…”

“…This can lead to poor model performance for some properties. Fine grinding can break up and increase the surface area of highly reflective materials such as metals, metalloids, and quartz, which can decrease absorbance increasing overall reflectance (see Figure 2a) (Deiss et al., 2019; Guillou et al., 2015). Cumulatively, these two effects produce less variable and better‐defined absorption features in MIR spectra, which can result in more accurate models with fine grinding.…”

“…Janik et al (2016) observed model improvements for clay and sand, and impairments for silt. Deiss et al (2019) used approximately 400 soils from the U.S. Midwest and tested the effect of fine grinding and replication on the performance of MIR models (trained by support vector machine) for five soil properties: clay, sand, pH, total organic C, and permanganate‐oxidizable C. Their results suggested that soil particle‐size reduction to <0.5 mm and collecting multiple spectra improve MIR predictions.…”

“…Janik et al (2016) observed model improvements for clay and sand, and impairments for silt. Deiss et al (2019) used approximately 400 soils from the U.S. Midwest and tested the effect of fine grinding and replication on the performance of MIR models (trained by support vector machine) for five soil properties: clay, sand,…”

Persistent soil carbon enhanced in Mollisols by well-managed grasslands but not annual grain or dairy forage cropping systems

“…We used anodized aluminum plates that fit 24 polystyrene sample cups (5.5-mL volume and 10-mm top opening diameter), each loaded with a subsample, dried for >48 h at 40 °C and 12 to 14% relative humidity. All measurements were conducted from 4,000 to 400 cm −1 at 4 cm −1 wavenumber resolution using 24 coadded scans ( 52 ). Spectra absorbance peaks were integrated using the local baseline technique, as described by Demyan et al.…”