The use of different interpolation methods and machine learning in the estimation of soil properties and mapping of spatial variability

Accurate determination of soil properties over a large area is crucial for the sustainable management and conservation of soils in line with their capabilities and for the selection and adaptation of precision agricultural techniques. In the first part of this study, the spatial prediction performances of Local Polynomial Interpolation (LPI) and Experimental Bayesian Kriging (EBK) were compared. In this study, the performances of Classification and Regression Trees (CART), Random Forest (RF) and Extreme gradient boosting (XGBoost) supervised machine learning algorithms in prediction of soil particle size distribution have been investigated. The prediction models were constructed by utilizing covariates such as remotely sensed data-derived plant and soil indexes, as well as topographic attributes. The Boruta approach was employed in variable selection. The reference data for training and evaluating the machine learning models consisted of particle size distribution results of 622 surface soil samples. Particle size distribution (sand, clay and silt), soil pH, lime content, electrical conductivity (EC) and cation exchange capacity (CEC) analyses were performed on 622 soil samples. Soil samples were collected from a 18,143 ha study area of in Ergani district of Diyarbakır in southeastern Turkey. In the first part of the study, the coefficient of variation between estimated and measured pH values in both models was quite similar. In the local polynomial interpolation method, the coefficient of variation of the estimated and measured data set in the lime content of soils was quite similar. The root mean square deviation (RMSE) values showed a decrease of 9.5%, 78.8% and 25.6% for pH, EC and CEC, respectively, in the EBK method compared to the LPI. The LPI-Lime method showed an RMSE improvement of 28.4% compared to EBK. The results showed that the error values of the estimations obtained with the EBK method were lower compared to the LPI method. On the other hand, the estimated pH and lime maps obtained with the LPI method were more compatible with the coefficient of variation of the sample dataset. In the second part of the study, the XGBoostClay model was the most accurate prediction machine learning model, with a R2 value of 0.74. The superiority was evidenced by a relative improvement of 21.36% in XGBoostClay RMSE compared to RFClay, and 44.5% compared to CARTClay. Likewise, the R2 values of models in the prediction of sand and silt content were 0.71 and 0.75 for the XGBoostSilt and XGBoostSand model. Among the covariates considered, Normalized Ratio Vegetation Index and Slope angle had the greatest impact on clay content (21%), followed by Topographic Position Index and Simple Ratio Clay Index (20%), while Terrain Ruggedness Index had the least impact (18%). The results suggested that computational techniques, such as Borutav, can be considered as a feasible option for choosing an adequate number of variables in digital soil mapping. This research is important to provide efficient spatial soil information for land management applications where rapid and accurate digital soil maps are required without the need for environmental variables.