JWSS - Journal of Water and Soil Science

Soil depth is a major soil characteristic commonly used in distributed hydrological modeling in order to present watershed subsurface attributes. It strongly affects water infiltration and accordingly runoff generation, subsurface moisture storage, vertical and lateral moisture movement, saturation thickness and plant root depth in the soil. The objective of this study is to develop a statistical model that predicts the spatial pattern of soil depth over the watershed from topographic and land cover variables derived from DEM and satellite image, respectively. A 10 m resolution DEM was prepared using 1:25000 topographic maps. Landsat8 imagery, OLI sensor (May 06, 2015) was used to derive different land cover attributes. Soil depth, topographic curvature, land use and vegetation characteristics were surveyed at 426 profiles within the four sub-watersheds. Box Cox transformations were used to normalize the measured soil depth and each explanatory variable. Random Forest prediction model was used to predict soil depth using the explanatory variables. The model was run using 336 data points in the calibration dataset with all 31 explanatory variables (18 variables from DEM and 13 variables from remote sensing image), and soil depth as the response of the model. Prediction errors were computed for validation data set. Testing dataset was done with the model soil depth values at testing locations (93 points). The Nash-Sutcliffe Efficiency coefficient (NSE) for testing data set was 0.689. The results showed that land use, Specific Catchment Area (SCA), NDVI, Aspect, Slope and PCA1 are the most important explanatory variables in predicting soil depth.

Land use/cover maps are the basic inputs for most of the environmental simulation models; hence, the accuracy of the maps derived from the classification of the satellite images reduces the uncertainty in modeling. The aim of this study was to assess the accuracy of the maps produced by machine learning based on classification methods (Random Forest and Support Vector Machine) and to compare them with a common classification method (Maximum Likelihood). For this purpose, the image of the OLI sensor of Landsat 8 for the study area (Sattarkhan Dam’s basin in the Eastern Azerbaijan) was used after the initial corrections. Five land uses including urban, irrigated and rain-fed agriculture, range and water body were considered. For conducting the supervised classification, ground truth data were used in two sets of educational (70% of the total) and test (30%) data. Accuracy indexes were used and the McNemar test was employed to show the significant statistical difference between the performances of the methods. The results indicates that the overall accuracy of Support Vector Machine, Random Forest, and Maximum Likelihood methods was 96.6, 90.8, and 90.8 %, respectively; also the Kappa coefficient for these methods was 0.93, 0.81 and 0.83, respectively. The existence of a significant statistical difference at the 95% confidence between the performances of the Support Vector Machine algorithm and the other two algorithms was confirmed by the McNemar test.

Digital soil mapping plays an important role in upgrading the knowledge of soil survey in line with the advances in the spatial data of infrastructure development. The main aim of this study was to provide a digital map of the soil family classes using the random forest (RF) models and boosting regression tree (BRT) in a semi-arid region of Ilam province. Environmental covariates were extracted from a digital elevation model with 30 m spatial resolution, using the SAGAGIS7.3 software. In this study area, 46 soil profiles were dug and sampled; after physico-chemical analysis, the soils were classified based on key to soil taxonomy (2014). In the studied area, three orders were recognized: Mollisols, Inceptisols, and Entisols. Based on the results of the environmental covariate data mining with variance inflation factor (VIF), some parameters including DEM, standard height and terrain ruggedness index were the most important variables. The best spatial prediction of soil classes belonged to Fine, carbonatic, thermic, Typic Haploxerolls. Also, the results showed that RF and BRT models had an overall accuracy and of 0.80, 0.64 and Kappa index 0.70, 0.55, respectively. Therefore, the RF method could serve as a reliable and accurate method to provide a reasonable prediction with a low sampling density.