Applicability of machine learning models for predicting soil organic carbon content and bulk density under different soil conditions

Abstract

A reliable overview of the spatial distribution of soil properties is a straightforward approach in soil policies and decision-making. Soil organic carbon (SOC) content, SOC stock and bulk density (BD) directly affect soil quality and fertility. Therefore, an accurate assessment of these crucial soil parameters is required. To do this, we used machine learning algorithms (MLAs) including, multiple linear regression (MLR), random forest (RF), artificial neural network (ANN), and support vector machine (SVM) with the help of environmental covariates to predict SOC content, BD, and SOC stock. The study was conducted in two different areas, Látókép and Westsik (East Hungary), both experimental research fields but different from physio geographic points of view. Thirty topsoils (0–10 cm) samples were collected for each study area using conditioned Latin Hypercube Sampling strategy. Environmental covariates were extracted from a digital elevation model (DEM) and satellite images based on the representation of soil forming factors. We validated the results by randomly splitting the dataset into a train (two-third) and test (one-third) and calculated the root mean square error and R2. Our results showed that RF provided the most accurate spatial prediction with R2 of about 80% for each soil property in both study areas. This study highlighted the importance of terrain attributes (including plan and profile curvature, elevation and valley depth) and NDVI derived from satellite images in presenting a spatial distribution of selected soil properties in two different areas. We conclude that comparing these methods can help to determine the most accurate maps under diverse geographical conditions and heterogeneities at different scales, which can be used in precision soil quality management.