Developing Near‐ and Mid‐Infrared Spectroscopy Analysis Methods for Rapid Assessment of Soil Quality in Illinois

Core Ideas Lithology‐based libraries may support reusable NIR models to predict soil quality. Monte Carlo feature selection and spectral refinement of NIR‐ and MIR‐based models. Use of randomly drawn validation datasets and overestimation of RPD. Estimation of plant growth via NDVI and relation to NIR‐ and MIR‐based models. Rapid, affordable assessment of soil quality (SQ) is needed to maintain soil health. We sought to build reusable near‐ (NIR) and mid‐infrared‐ (MIR) based models using a regional library of accepted SQ indicators and Normalized Difference Vegetation Index (NDVI). This described soil health in 228 topsoil samples from 57 grain fields in Illinois paired to compare the influence of conventional, organic or conservation‐tillage practices. Predictive models for all SQ indicators were developed with Partial Least Squares Regression (PLSR) using: (i) whole NIR or MIR spectra, (ii) refined spectra that included features associated with organic functional groups, or (iii) refined spectra that were identified as important using Monte Carlo feature selection (MCFS). Comparison of model building methods using randomly or systematically divided calibration and validation sets suggests that the random method overestimated model Residual Prediction Deviation (RPD). Use of systematically selected validation sets produced acceptable (RPDs ≥ 1.4) models for soil organic C (SOC), total N (TN), and exchangeable Ca and Mg using whole NIR spectra. Use of spectral refinement techniques improved MIR‐based models for SOC, TN, and exchangeable Ca to RPDs ≥ 1.4. Poor performance of models predicting other SQ indicators, including particulate organic matter (POM), potentially mineralizable N (PMN), and fluorescein diacetate hydrolysis (FDA), is attributed to the influence of mineralogy associated with different types of loess. Use of lithology‐based libraries may support reusable and affordable NIR‐based models to predict SOC, TN, Ca and Mg. The addition of weather or moisture co‐variates may be needed to build models reliant on NDVI.