MidDRIFTS-PLSR-based quantification of physico-chemical soil properties across two agroecological zones in Southwest Germany: generic independent validation surpasses region specific cross-validation

Abstract

Development of spectroscopic prediction models via partial least squares regression (PLSR) suggests that model performance is highly affected by means of calibration and nature of the dataset. This study compares the predictive performance of PLSR models obtained by cross-validation and independent validation to quantify physico-chemical soil properties from their mid-infrared diffuse reflectance Fourier transform spectra (midDRIFTS) across two contrasting regions, Kraichgau (K) and Swabian Alb (SA), in Southwest Germany. We evaluated the capability of midDRIFTS-PLSR models for predicting total carbon (TC), organic carbon (TOC), inorganic carbon (TIC), nitrogen (TN), mineral N (N_min), C:N ratio, hot water extractable C and N (C_HWE, N_HWE), microbial biomass C and N (C_mic, N_mic), pH, bulk density, and clay, silt and sand contents of 126 soil samples. Based on calibrated models, most soil properties were predicted successfully using either calibration approach with residual prediction deviations ≥3 and R² > 0.9, except for N_min, C/N ratio, pH, bulk density and sand. However, predictive performance of generic independent validation derived models (GIC) of test set was considerably higher than generic cross-validation models. Validation using GIC models gave relatively the same predictive performance with those obtained in calibration except for N_min. Validation of region specific cross-validated models, however, resulted in successful predictions only for TC, TIC, TOC and TN in SA and TC and TIC and TOC in K. Our results show the superiority of independent validation over both generic and region specific cross-validation as a robust tool for predicting soil properties without further laboratory measurements.