How well can VNIR spectroscopy distinguish soil classes?

Visible near-infrared (VNIR) spectra can provide rich information on soil physical and chemical properties, which implies the possibility of using soil spectra to aid in the discrimination of soil types. Pedological soil classification systems use a selected set of soil properties to identify diagnostic horizons and features, and to build a classification key. This research explored the application of VNIR spectra to classify typical soil profiles collected in Anhui province, China. The 279 soil profiles used are classified into five orders (Cambosols, Vertosols, Argosols, Primosols and Anthrosols), six suborders and 21 groups according to Chinese Soil Taxonomy. Soil spectra were collected within 350-2500 nm and principal component analysis (PCA) was applied to reduce data dimension. These principal components were used as independent variables in multinomial logistic regression for soil classification. Topsoil spectra, subsoil spectra and their combination were compared for prediction accuracy. Accuracy achieved at the level of suborder using spectra of topsoil, subsoil and combined horizons were 76.3%, 71.3% and 70.3% respectively, while the results for the level of soil group using the topsoil horizon was 40.5%. Since topsoil spectra alone achieved a prediction accuracy of more than 75%, reflectance spectroscopy can be judged a promising tool for soil classification. Taxonomic distances between classes calculated on the basis of physio-chemical properties and spectra were quite different, showing that the concept of distance between classes in feature space depends on the features chosen for evaluation. Taxonomic distances can serve as a supplement for better selection and evaluation of prediction models.