Remote sensing and machine learning methods are gaining popularity in studies of the spatial distribution of soil indicators. The use of approaches based on these methods improves the accuracy of digital maps of soil properties. Using the example of a field with developed erosion located in the chernozem zone of the Republic of Tatarstan (Russia), we studied the use of remote sensing and machine learning to obtain, on a scale of one field, digital maps of the organic carbon content, available forms of nitrogen, phosphorus and potassium, silt and clay fractions for precision farming purposes. Spectral parameters of bare soil obtained from the Landsat 8 OLI and Sentinel 2 satellites were used as predictors. Linear models (MLR), support vector regression (SVM) and random forest (RF) were used as models. Model performance was assessed based on RMSE, MAE and R² after bootstrap. It is shown that the SVM and RF models outperform the MLR models. The best were the RF models for organic carbon (R² = 0.91), available nitrogen (R² = 0.83), potassium (R² = 0.81), and clay (R² = 0.67). SVM models were more accurate for available phosphorus (R² = 0.91) and silt content (R² = 0.93).

This work was supported in part by the Russian Foundation for Basic Research, research project № 19-29-05061-mk