Journal article
Pitfalls in the use of middle-infrared spectroscopy: representativeness and ranking criteria for the estimation of soil properties



Publication Details
Authors:
Ludwig, B.; Linsler, D.; Hoeper, H.; Schmidt, H.; Piepho, H.; Vohland, M.
Publisher:
ELSEVIER SCIENCE BV
Publication year:
2016
Journal:
Geoderma
Pages range:
165-175
Volume number:
268
Start page:
165
End page:
175
Number of pages:
11
ISSN:
0016-7061

Abstract
Middle-infrared spectroscopy (MIRS) is an established method for estimating the contents of soil organic carbon (SOC) and total soil nitrogen (N). However, obtained estimation accuracies vary between studies and only few studies are available that deal with C and N fractions. Objectives were to determine estimation accuracies for contents of SOC, microbial biomass C (C-mic) and C and N fractions for two samples of surface soils using different software packages (with different data treatments) and to discuss the usefulness and limitations of MIRS for a quantitative assessment of soil properties. Eighty-four surface soils were collected from arable sites from eight German states; their middle infrared spectra were recorded and their physical, chemical and biological properties determined. Estimates of SOC contents were obtained with WinISI software in cross-validations with and without removal of spectral (H>10) outliers and units with large deviations between measured and estimated values (T>2.5). Sample I (all 84 soils) consisted of soils from different horizons (partly with a substantial fraction of tangle of roots) and comprised of pseudo-replicates (different managements, but same mineralogy); for this ill-defined sample WinISI achieved an apparently excellent estimation accuracy when suspected outliers were removed. We suggest that T outliers should not be removed from samples in soil infrared studies except for preliminary evaluations. In contrast, for the consistently defined subset sample II (i.e., soils were taken from Ap and M-Ap horizons from 51 German arable sites with typical SOC contents and without pseudo-replicates) without outliers only a good estimation accuracy was reached. This indicates that besides a search for optimum estimation accuracies, equal attention should be paid to the representativeness of the sample for a specific population, an appropriate handling of suspected outliers and the generalizability of the MIRS results. With respect to accuracies, we obtained good results, approximative quantitative results or accuracies with the potential to discriminate between high and low values for all C and N fractions (except for light-fraction N) and also C-mic. An estimation of these properties without infrared data using the contents of SOC, N, pH, sand, silt and clay in multiple linear regressions was generally slightly less successful than the MIRS estimates using OPUS. However, when we created artificial spectra based solely on the measured pH, contents of SOC and N and texture data without any real underlying infrared data - and then used them for a PLS regression in OPUS, the performance was similar to the MIRS estimates, with a slight difference for passive C and N. Overall, our study indicates that MIRS is a useful method for an estimation of the spectrally active main constituents SOC and N and possibly for passive C and N. However, there is not much benefit of using MIRS to obtain a spectral assessment for those properties, where approaches without infrared data (either multiple linear or PLS regressions using pH, SOC, N and texture data) give estimates of similar accuracy, which was the case for of C-mic, light-fraction C and N, mineral-associated C and N or intermediate C and N for the dataset investigated here. (C) 2016 Elsevier B.V. All rights reserved.


Keywords
Calibration set selection, carbon fractions, Multiple linear regression, nitrogen fractions, Outlier detection, Partial least squares regression


Authors/Editors

Last updated on 2018-14-11 at 16:03