Comparison of Different Data Mining Methods for Digital Mapping of Soil Particle-size Fractions in Lands of Semnan Plain

Knowledge about the spatial distribution of particle-size fractions in different areas is required for various land management applications and resources, modeling, and monitoring practices. In recent years, with the advancement of data mining methods and the availability of cheap data from satellite imagery, digital soil mapping methods have been developed to predict the spatial distribution of primary soil particles. The objective of this study was to conduct a spatial prediction of particle-size fractionssuch as clay, sand and silt using digital soil mapping in agricultural lands in Semnan. To achieve these goals, a total of 84 soil samples were collected from 0 to 20 cm of soil surface. Also, the environmental variables were obtained from OLI Satellite Landsat to make dependence with soil particles. In this study a linear model such as Partial Least Squares (PLS) and two non-linear models, including Random Forest (RF) and Stochastic Gradient Boosting Machin (GBM) were used for spatial prediction of particle-size fractions. The models were calibrated and validated by the 10-fold cross-validation methods. Three statistics, such as Root Mean Squared Error (RMSE), Coefficient of determination (R2), and Mean Absolute Error (MAE) were used to determine the performance of the investigated models. Values of RMSE, R2, and MAE statics of RF model for prediction of sand, silt and clay were (15.6, 0.35, 12.62), (11.49, 0.33, and 9.34), and (8.42, 0.28, and 5.9), respectively. These results indicated that the most accurate model for the prediction of particle-size fractions was the RF model. Also, the results showed that the most important environmental covariates for predicting particle-size fractions were band 10 (B10), band 5 (B5), and the gypsum index (GI). This indicated that the variables containing the near-infrared and infrared thermal waves had a major contribution to explaining the spatial patterns of particle-size fractions.