Diffuse reflectance mid-infrared spectroscopy is viable without fine milling

Abstract

While diffuse reflectance Fourier transform mid-infrared spectroscopy (mid-DRIFTS) has been established as a viable low-cost surrogate for traditional soil analyses, the assumed need for fine milling of soil samples prior to analysis is constraining the commercial appeal of this technology. Here, we reevaluate this assumption using a set of 2380 soil samples collected across North American agricultural soils. Cross-validation indicated that the best preprocessing (standard normal variate) and model form (memory-based learning) resulted in very good and nearly identical predictions for the <2 mm preparation and fine-milled preparation of these soils for total organic carbon (TOC), clay, sand, pH and bulk density (BD). Application of larger models built from the USDA NRCS mid-DRIFTS library also resulted in minimal performance differences between the two sample preps. Lower predictive performance of the existing library was attributed to less-than-perfect spectral representativeness of the library. Regardless of model form, there was very little variability between replicates of the <2 mm prep, suggesting that the lack of fine milling did not lead to more heterogeneous subsamples. Additionally, there was no relationship between residual error and soil texture, implying these results should be robust across most soil types. Overall, in agreement with other recent findings, these results suggest that routine scanning of standard <2 mm preparation does not degrade predictive performance of mid-DRIFTS-based inference systems. With good standard operating procedures including quality control and traditional analysis on a small percent of samples, mid-DRIFTS can become a routine tool in commercial soil laboratories.