Application of machine learning models and Euclidean distance to predict soil spatial properties

Accurate estimation of the spatial properties of forest soil is essential for sustainable land management. This research aimed to map key soil properties in temperate forests using hybrid machine learning models. The integration of predictive and geostatistical techniques has gained prominence in soil science, with hybrid approaches enhancing prediction accuracy. Precise soil maps serve as a foundational resource for effective soil management, motivating the development of more accurate and cost-efficient mapping methods. We used hybrid machine-learning techniques that incorporated Euclidean distance (Dis), remote sensing (RS) data, and digital elevation models (DEM) to improve spatial predictions. We hypothesized that integrating Euclidean distance data into predictive models would boost the accuracy of soil property maps. Model performance was evaluated using root-mean-square error (RMSE), coefficient of determination (R²), mean absolute error (MAE), and concordance correlation coefficient (CCC). The hybrid RF+GA+Dis model offered the highest accuracy for predicting soil organic carbon (RMSE = 1.810, R2 = 0.763, MAE = 1.883, CCC = 0.803), with similar improvements observed for soil nitrogen, organic carbon stock, bulk density, calcium carbonate, and soil texture. Hybrid models consistently demonstrated a superior performance over individual machine learning methods such as Random Forest (RF) and Genetic Algorithm (GA). Incorporating ancillary data, especially from DEM and RS, substantially ameliorated prediction accuracy for soil physicochemical properties. Among the tested models, those using a broader range of covariates demonstrated a superior performance, with RF+GA+Dis outperforming RF+Dis. These findings confirm that integrating machine learning with comprehensive spatial data is a cost-effective and reliable approach for generating high-resolution soil maps, facilitating precision land management and informed decision-making. This highlights the value of hybrid modeling and diverse covariates in advancing soil property prediction.