Spatial Prediction Some of the Surface Soil Properties Using Interpolation and Machine Learning Models

Accurate and detailed spatial soil information over the landscape is essential for the precision monitoring of land resources, hydrological applications, land use management. The present study aimed to predict the spatial prediction of SOC, CCE, Clay, Silt, and Sand in the Qorve-Dehgolan region, Kurdestan province. Qorve-Dehgolan region, with mean annual temperature and precipitation of 12 C° and 348 mm (20-year statistical period), has soil moisture and temperature regimes xeric and thermic, respectively. A total of 145 samples were collected from the topsoil (0-30 cm) based on a random sampling pattern. Then, all of the soil samples were transferred to a soil laboratory for physicochemical analysis. Random forest (RF) as a nonparametric model and Ordinary kriging (OK) and inverse distance weighting (IDW) as an interpolation method were used for modeling the soil properties and their spatial autocorrelation. All steps of modeling for RF and interpolation methods (OK and IDW) were performed in RStudio, ArcGIS and, GS+ software, respectively. A total of 30 environmental covariates, including the Digital Elevation Model (DEM) derivatives in the SAGA GIS 7.3 software and Landsat 8 satellite reflective band data in the ERDAS IMAGINE software, were developed as environmental variables. All of the environmental covariates were resampled at resolution-30 m. The most appropriate covariates were selected according to the variance inflation factor (VIF). Modeling of soil properties was performed according to 80% and 20% of data sets, respectively for calibration and validation, and two statistics of root mean square error (RMSE) and determination coefficient (R2) was used to determine the accuracy of the models.

Seven variables including SAVI, EVI1, GNDVI, RVI1, DEM, Channel Network, and TPI were selected from the 30 variables prepared as the most appropriate auxiliary variables based on the variance inflation index. Four remote sensing variables include the adjusted soil vegetation index (SAVI), the greenness of the normalized differential vegetative index (GNDVI), the relative vegetation index (RVI) and the Enhanced vegetation index (EVI), and three geomorphometric attributes including, digital elevation model (DEM), Vertical distance to channel network and the topographic position index (TPI) were the most important parameters. The results of modeling showed that RF model for soil organic carbon variable (R2 = 0.5 and %RMSE= 0.4), calcium carbonate equivalent (R2 = 0.4 and %RMSE = 11.61), clay variable (R2= 0.21 and %RMSE=5.65), the Silt variable (R2 = 0.15 and %RMSE= 7.24) and, Ordinary kriging methods for sand variables with (R2 = 0.14 and %RMSE = 10.26) was the most accurate than RF and IDW models. Among the semi-variogram models, the exponential model had the best performance for soil organic carbon, clay, silt, and sand percentage, with the exception of CCE which follows the spherical model. The results of spatial autocorrelation showed that for both variables CCE and Sand had a strong class and, the rest had a moderate class. The highest values of the semi-variogram sill were related to the calcium carbonate equivalent and clay, and the lowest values were related to the soil organic carbon and sand contents. These results indicate that, the existence of a random pattern or weak spatial structure in the samples that used to calculate the experimental semi-variogram. Among the seven environmental covariates were used for spatial modeling of top-soil organic carbon, calcium equivalent carbonate, and clay, the geomorphometric attributes such as digital height model, topographic position index and vertical distance to channel network are of the most important and NDVI, SAVI and RVI covariates were more important in predicting of sand and silt properties.