Incorporating environmental variables, remote and proximal sensing data for digital soil mapping of USDA soil great groups

Abstract

ABSTRACT The present study was conducted to evaluate the effectiveness of combining proximal, and remote sensing with environmental variables for predicting USDA (United States Department of Agriculture) soil great groups (the third hierarchical level of USDA soil classification system) in a semi-arid region located at Jouneqan district in Chaharmahal & Bakhtiary province, Iran. Accordingly, two predictive models, including support vector machine (SVM), and multinomial logistic regression (MLR) using different covariates, including lithology, geomorphology, remote sensing-derived vegetation indices, DEM (digital elevation model)-derived attributes, diffuse reflectance spectroscopy-derived soil colour qualifiers, and magnetic susceptibility were examined. A total of 102 soil profiles were excavated, described, and classified up to the great group level, and soil samples were collected from various genetic horizons. The cross validation leave-one-out (LOOCV) was used as validation approach, and the performance of the models was assessed using the kappa coefficient (κ), and overall accuracy. Results showed that considering the κ values (κ = 0.6–0.8), the classification performance identified as substantial for both MLR, and SVM when geospatial data, soil colour qualifiers, and magnetic susceptibility were used together as predictors. However, the MLR classifier outperformed SVM (κ: 0.78, and 0.66, respectively). Chroma, normalized difference vegetation index (NDVI), soil adjusted vegetation index (SAVI), geomorphology map, and some terrain attributes, including slope were among the most important predictors. Our findings confirmed that combining geospatial data, and proximal sensing information in multivariate classification algorithms could improve soil classification accuracy in a quick, and cost-efficient way.