Digital Mapping of the Humus Horizon Thickness in Soils of the Cis-Salair Plain Using the Random Forest Machine Learning Algorithm

Results of digital mapping of the humus horizon thickness (HHT) in the soils of the Cis-Salair Plain using the Random Forest (RF) machine learning algorithm implemented on the Google Earth Engine cloud online platform are reported. A total of 92 predictors are employed to characterize the soil formation factors, including climate, relief, vegetation, spatial position, and soil properties. Training (n = 718) and validation (n = 130) datasets are constructed based on the archive materials (1974–1984) of ZapSibNIIgiprozem (Western Siberian Research, Design, and Survey Institute for Land Use Planning). The following indicators of the HHT modeling efficacy using the RF algorithm are obtained: coefficient of determination for training dataset $$R_{{{\text{TD}}}}^{2}$$ = 0.88; coefficient of determination for validation dataset $$R_{{{\text{VD}}}}^{2}$$ = 0.12; root mean square error RMSEVD = 9.7 cm; mean absolute percentage error MAPEVD = 24.3%; and mean absolute error MAEVD = 6.5 cm. The modeling accuracy estimated with MAPEVD is satisfactory. Actual data show that HHT varies from 3 to 110 cm with the trend of a decrease from northwest to southeast. The lowest (3 cm) average HHT values are typical of meadow–chernozemic solonetz (Solonetz (Salic)) and the highest (61 cm), of ordinary meadow soils (Mollic Gleysols).