Machine Learning Approach for Soil Organic Carbon Prediction Using Auxiliary Environmental Variables in Agricultural Lands

Abstract The traditional method for mapping soil properties relies on field surveys, with the spatial variability of the data being influenced by the study scale and requiring the expertise of experienced professionals. Therefore, utilizing new data mining models to make digital maps of soil properties is essential. Moreover, modeling and mapping the spatial distribution (SPD) of plant nutrients in soil is crucial for enhancing agricultural productivity and achieving sustainable development. The present research aimed to digital mapping of soil organic carbon (SOC) using two data mining techniques namely, Cubist (Cu) and Random Forest (RF) models in some agricultural lands of the Fars Province, Iran. Accordingly, for this survey, 270 soil samples were gathered in a systematic manner of topsoil (0–0.3 m), and the percentage of SOC was measured. Nineteen ancillary variables derived from Landsat8 images, and the Digital Elevation Model (DEM) were used to predict this parameter. The results indicated that, according to ten-fold cross-validation both the RF and Cu models provided similar accuracy for predicting SOC, with RMSE values of 0.35 and 0.36 and MAE values of 0.25 for both models. Also, the finding emphasized that the most crucial ancillary variables that significantly influence the prediction of spatial variation in SOC were CNBL, MRVBF, and NDVI. Thus, given the understanding of the region and the variability in SOC, the RF model can be considered an appropriate method for mapping soil properties, particularly SOC.