Diffuse reflectance spectroscopy (DRS) has shown its potential as a feasible, rapid and non-invasive soil characterization tool. Nevertheless, the use of whole VisNIR spectra in DRS models often incorporates different disruptive and masking effects, eventually producing inefficient model predictions. Thus the careful choice of informative spectral variables is a significant step toward producing robust and useful DRS-based models. This study evaluated the feasibility of combining variable indicator-based DRS outputs and geostatistical interpolations to rapidly produce soil spatial variability maps for six soil properties [sand, clay, silt, total carbon (TC), total nitrogen (TN) and loss-on-ignition organic matter (LOI)]. A total of 300 samples were collected from three catenas of Transylvanian Plain, Romania. First derivative spectra were used to calculate Pearson's correlation coefficient (r), biweight midcorrelation (bicor), mutual information based adjacency (AMI), variable importance in the projection (VIP), and their combinations. This variable indicator suite was combined with an ordered predictor selection (OPS) method to choose the optimum number of spectral variables (NSV). This method was tested with partial least squares regression (PLSR) and support vector regression (SVR) with independent validation. Results indicated that the variable indicator-based SVR model yielded superior predictability relative to full-spectrum PLSR model for all soil parameters. Moreover, both PLSR and SVR optimal models used the identical best variable indicators. While AMI appeared as the best indicator for four soil attributes (clay, TN, TC and LOI), bicor was selected as the best indicator for sand and silt. Spatial variability mapping using optimal SVR model outputs satisfactorily demonstrated management and landscape dynamics across the catenas like the place of manure stockpiles. Summarily, the results of this study indicated that a successful combination of OPS-based variable indicators and their subsequent incorporation into DRS-based chemometric models can potentially improve model predictions that can be further combined with geostatistical interpolation methods to produce spatial maps of soil properties.
- Diffuse reflectance spectroscopy
- Ordered predictor selection
- Spatial variability
- Variable indicator
- Visible near infrared