Infrared Spectroscopic Predictions of Soil Properties for Agricultural Monitoring: Statistical and Soil Spectroscopic Considerations

Background : Pseudo‐independent validation (PIV) using regional datasets is common in visible‐to‐near‐infrared (vis–NIR) and mid‐infrared (MIR) spectroscopy studies and produces accurate predictions but is relevant only to regional‐scale monitoring of known sites. However, site‐specific evaluations with true independent validations (TIV) are essential for real‐world applications. Aim : Objectives were to compare results of regional PIV versus site‐specific TIV for soils from five agricultural experiments. Methods : Vis–NIR and MIR spectra of 360 soils were recorded and regressions for total C (C t ) and N, organic C, sand, clay, pH, and cation exchange capacity were calculated with regional PIV and TIV for new sites using small ( n = 20) or large ( n = 52) local calibrations or with the remaining sites as a library (SSL) without or with spiking. Results : For regional PIV, optimal predictions ( R 2 ≥ 0.85) for all properties were achieved with MIR or fused spectra with support vector machine regression. For site‐specific TIV, accuracies varied across properties and sites, and C t predictions using MIR were significantly better using a spiked SSL than with a small local calibration. However, for other properties, there was no benefit of a spiked SSL over local calibration. For C t , normalized root mean square errors indicated which sites were well‐represented by spiked SSLs, but results were not related to differences in soil type, parent material, texture, or spectral principal components. Conclusions : Aims and accuracies differ largely between regional PIV and site‐specific TIV. For site‐specific TIV, successful predictions of C t using spiked SSLs do not necessarily translate to other properties.