Controlling factors in the variability of soil magnetic measures by machine learning and variable importance analysis

Abstract

The present work is the first effort to apply machine learning approaches and variable importance analysis (VIA) to the determination of environmental attributes that regulate the variation of magnetic susceptibility (MS) measurements (χlf, χhf, and χfd) in landscape scale. We depicted a total of 346 soil samples (0–20 cm depth) by stratified random sampling in the west of Iran. MS measurements and some soil properties comprising clay, silt, sand, electrical conductivity (EC), soil organic carbon (SOC), pH, and calcium carbonate equivalent (CCE) were measured. Fifty environmental variables from different sources (soil analyses, topographic attributes, remote sensing data, and thematic maps) and three machine learning approaches (Support vector machine, SVM), Random forest, RF), and Cubist) were used for modeling. The results revealed that the RF model showed the highest accurateness modeling χhf (R2 = 0.49, RMSE = 0.37) and χfd (R2 = 0.40, RMSE = 0.43). Also, the Cubist model presented acceptable accurateness in the modeling of χlf (R2 = 0.50, RMSE = 0.37). Among the independent variables used in the modeling of MS attributes, the distance from mines was identified as the most important variable, due to the release of ferrimagnetic minerals to the nearby environment. In the next rankings, soil properties affected by parent materials (silt and sand), as well as SOC and CCE “as the diamagnetic minerals” showed great contribution to the variability of MS. Moreover, some indices derived from remote sensing and topographic data including Ferrous Iron, Ferrous mineral, brightness index (BI), and Valley depth were important. This research confirmed the high potential of machine learning and VIA to find out the factors that control magnetic susceptibility across the study area. On the other hand, more advanced technologies are still needed to differentiate the contribution of each controlling variable in the detailed scales.