Spatial prediction of soil properties using random forest, k-nearest neighbors and cubist approaches in the foothills of the Ural Mountains, Russia

Soil organic matter (SOM) and pH are important indicators to evaluate soil quality. This study applied several machine learning (ML) techniques, namely Random Forest (RF), k-Nearest Neighbor method (kNN) and Cubist, to digital map SOM and pH (H2O and KCl) contents in the foothills of the Ural Mountains, Russia. For this purpose, a total of 52 soil samples were collected from the topsoil depth (0-20 cm). The environmental variables were derived from a digital elevation model (DEM), satellite imagery (Sentinel-2A) and land use/land cover (LULC) map. The ML models were calibrated and validated by the leave-one-out cross-validation approach. The coefficient of determination (R-2) and root-mean-square error (RMSE) were used to determine the ML model performance. According to the R-2 and RMSE metrics, Cubist method resulted in the most accurate spatial prediction for SOM (R-2 = 0.64 and RMSE = 1.95), while RF approach showed the highest performance to predict pH H2O and KCl (R-2 = 0.49; RMSE = 0.45 and R-2 = 0.44; RMSE = 0.61, respectively). Results showed that remote sensing data were the key variables to explain the variability of all soil properties. Sentinel-2A bands B8A, B7 and B8 were the most effective covariates in predicting SOM, whereas spectral indices GNDVI and SAVI explained most of the spatial distribution of both soil pH. According to the generated maps, the highest SOM concentrations (4-8%) were found under the forest and especially at the bottom of the slopes, which is consistent with favorable conditions of organic carbon accumulation and its redeposition under the influence of water erosion. More acidic soils (pH H2O < 6) were also located under the forest, consisting of mixed coniferous broad-leaved species, which also affects the soil acidity. This study confirms the necessity to use various ML techniques to predict individual soil properties. Overall, our findings and generated maps can provide useful information for future digital soil mapping of areas in similar climatic and geographic conditions.