Uncertainty Assessment of Soil Property Estimations and Spatial Soil Models Using High-Resolution VIS-NIR (diffuse reflectance) Spectra.

Visible-near-infrared diffuse reflectance spectroscopy has been incorporated into soil science for over 20 years as means of rapid prediction of soil properties. Soil spectral data is seen as representative of the entire soil and offers more information than conventional soil analyses since it can be used to infer on multiple soil properties concomitantly. But rarely has it been incorporated into the spatial modeling of soil properties. We assert that by including spectral data into the spatial modeling process, the predictions of target soil properties can be improved. This was quantified using aspatial and spatially explicit uncertainty assessment.

Our objectives for this study were to (i) assess the accuracy, bias and uncertainty of soil property spectral prediction models, (ii) assess the uncertainty of spatial predictions of the measured soil properties, and (iii) assess the uncertainty of soil property spatial predictions using soil spectral data. The assessments were performed on soil data acquired from an agricultural village located in southern India. Over 250 point soil samples were collected with geographic coordinates, analyzed for critical properties related to soil fertility, and scanned in the visible-near-infrared spectral region. Partial least squares regression was applied to the soil spectra to predict soil property concentrations and uncertainty analysis was performed with jackknifing. The uncertainty of the measured soil property spatial predictions was determined using Bayesian kriging with and without environmental covariates. Bayesian kriging was also applied to the spatial modeling of soil properties with the inclusion of soil spectral data to determine the overall uncertainty of the combined model. Our results indicate an overall improvement in the uncertainty of the spatial soil property models by incorporating soil spectral data, although some properties observed greater reductions in uncertainty than others.