Rapid Assessment of Soil Quality in Illinois Using Near- and Mid-Infrared Spectroscopy.

Abstract: Interest in use of near- (NIR) and mid-infrared (MIR) spectroscopic methods for soil quality assessment has grown rapidly.  Ideally robust calibration of models can be developed to rapidly and affordably predict soil quality indicators (SQIs) that are more costly to measure. This study investigated 468 topsoil samples collected from Illinois grain farms located in 5 soil regions that were using conventional tillage (CT), conservation-tillage (NT) or organic practices. Partial least squares regression (PLSR) and random forest (RF) algorithms were used to predict SQIs, including soil organic carbon (SOC), total N (TN), soil C and N ratio (C: N), soil pH, particulate organic matter (POM), potentially mineralizable nitrogen (PMN), fluorescein diacetate hydrolysis (FDA), and soil nutrient (P, K, Ca, Mg, Fe) abundance from the whole NIR or MIR spectra and spectral features explanatory for organic functional groups. Monte Carlo feature selection (MCFS) process was used as a variable selection tool for PLSR models. In our study, the NIR models outperformed MIR models slightly, and both NIR and MIR methods were better able to predict SOC, Ca, and TN than other SQIs.  Model performance was not improved when spectral ranges associated with organic functional groups were primarily used, while variable selection tool improved model performance in MIR regions. Variables selected by RF explained features of soil labile components better than PLS regression, which led to slightly better performance of RF calibration models. We also explored the relationship between SQIs and crop productivity using normalized difference vegetation index (NDVI), and found that selected SQIs could be used to predict NDVI better than using the whole sets of SQIs. This work demonstrates how NIR and MIR spectroscopy can be used to quantify soil quality on a regional scale, and noted the importance of proper interpretation of prediction models, not only statistically but also from their physical meanings.

Key Words: Soil quality indicators (SQIs); Near- (NIR) and mid-infrared spectroscopy (MIR); Partial least squares regression (PLSR); Random forest (RF); normalized difference vegetation index (NDVI); Monte Carlo feature selection process (MCFS)