Large-Scale Soil Mapping Using CART Algorithm: Advantages and Restrictions

Abstract: This manuscript presents the results of testing CART (Classification and Regression Trees) algorithm as the basis for large-scale soil mapping. The study was conducted at the test site with agricultural fields located in Liski district of Voronezh oblast, Russia. The soil cover included chernozems and meadow-chernozemic soils. Modeling was carried out for the distinguished soil taxa (soil names) and for six soil properties: thickness of the A1 horizon, thickness of the humus layer (A1 + AB horizons), depth of effervescence, sand content, silt content, and humus content. The brightness of spectral bands (R, G, B), slope, aspect, absolute height (m a.s.l.), and standard deviation of the absolute height in a point (pixel) were used as explanatory variables. The quality of decision trees was assessed on the basis of the cross-validation error value. For the thickness of the A1 horizon and for the contents of humus and silt, the model based on satellite data only turned out to be better. For the total thickness of the A1 and AB horizons, the depth of effervescence, and the sand content, more accurate results were provided by the model based on the satellite data (brightness of spectral bands) and morphometric parameters of the relief. Among the considered predictors, the brightness of the blue band and absolute height were most important. The capacity to work with abnormally distributed input data appears to be the main advantage of CART algorithm. The method applied in the study can be used to construct soil maps, especially at the initial stage of the assessment of agricultural potential of the territory or as an auxiliary tool upon working with a complex soil cover. © Pleiades Publishing, Ltd. 2026.