In this paper we investigated the possibility of using the VisNIR spectra of dry ground soils to predict properties of soils scanned in their natural physical state with unknown water content. We regard this as an important nexus between large soil spectral libraries (based on dried and ground soil samples) and in-field use of VisNIR where variable soil moisture and secondary structure affect predictions of soil properties of interest. External parameter orthogonalization - EPO - was used in developing a partial least squares regression (PLSR) model to predict clay and organic C contents of soil samples with variable moisture contents. The Texas Soil Spectral Library based on spectra of dried and ground soil samples along with spectra of intact and dried and ground soil cores from Central Texas USA were used for EPO-PLSR model calibration and validation.Using EPO, three matrices were developed to project VisNIR reflectance spectra to a subspace insensitive to soil moisture. The first matrix P1 was developed using spectra collected from rewetted samples that had been dried and ground. The second and third matrices, P2 and P3 were developed using spectra from soil core samples that were in their natural physical state when scanned and either field-moist (P2), or air-dried (P3). The results showed that EPO-PLSR successfully removed the effect of moisture without knowing the moisture at time of scanning and substantially improved the prediction of clay content compared to organic C content. For clay content, the validation results were as follows: No correction, R2=0.63, RMSEP=355gkg-1; P1 correction, R2=0.73, RMSEP=141gkg-1; and P2 correction, R2=0.77, RMSEP=90gkg-1. For organic C content, the validation statistics were: No correction, R2=0.49, RMSEP=9.4gkg-1; P1 correction, R2=0.51, RMSEP=7.5gkg-1; and P2 correction, R2=0.53, RMSEP=7.3gkg-1. Corrections of soil intactness alone had the following results for clay content: No correction RMSEP=125gkg-1 and P3 correction RMSEP=97gkg-1, and for organic C content: No correction RMSEP=7.5gkg-1, and P3 correction RMSEP=7.4gkg-1. Model results using the P2 matrix were consistently better than using P1. Particularly in predicting clay, P2 reduced the bias, non-unity, and lack of correlation, possibly because P2 accounted for the effect of natural aggregation of the soil in addition to soil moisture. Improvements for correction for intactness alone were small in clay and not significant in organic C content. We concluded that it is feasible to apply the EPO algorithm to employ VisNIR models from dried ground spectral libraries for prediction of soil properties based on field scans of soils in the natural physical state and at variable water contents.
- External parameter orthogonalization
- Organic carbon
- Partial least squares regression
- Proximal sensing
- Soil spectral library
ASJC Scopus subject areas
- Soil Science