Determination of possible soil boundaries and mapping of some soil properties in lands located on different physiographic units using digital soil mapping approaches

This study aims to delineate potential soil boundaries and map the spatial distributions of certain soil properties (e.g., soil organic carbon (SOC), soil inorganic carbon (SIC), and textural fractions) using digital soil mapping approaches in lands located within different physiographic units. The locations within the study area were determined using the conditional Latin hypercube sampling (c-LHS) method. A total of 214 samples were collected from 107 selected points at both the surface (0–30 cm) and subsurface (30–60 cm) depths. The reflectance values of the sampling points were measured in the laboratory using visible-near infrared spectroscopy (vis-NIR), and their accuracies were subsequently verified. Analyses for SOC, SIC, texture, pH, and electrical conductivity (EC) were performed on the sampled points. The obtained data, along with various environmental variables (such as geomorphometric, climatic, remote sensing, and land use data), were evaluated using regression-based (Regression Kriging, Random Forest, and Support Vector Regression) as well as classification-based machine learning algorithms to produce digital prediction maps. Furthermore, the DSMART algorithm was employed to resample soil mapping units and conduct uncertainty analyses, with model performance assessed through quantitative error measurements. The study demonstrates that different machine learning models achieve varying levels of success in predicting the depth and spatial distributions of soil properties, with the advantages and limitations of each model discussed in detail. The results provide significant insights for precision agriculture applications, land management, and sustainable soil use, contributing to a better understanding of the interactions between soil formation processes and environmental factors. Overall, this study offers important methodological and practical contributions to the field of digital soil mapping, showcasing the effectiveness of modern machine learning techniques in the spatial modeling of soil data. Keywords: Digital soil mapping, Regression based models, Classification based models, DSMART, Vis-NIR spectroscopy