Integrating proximal sensing data (soil spectra and magnetic susceptibility) with the common covariates to predict soil carbon pools in semiarid regions

Abstract

Accurate mapping of soil carbon stocks in semiarid and arid areas can be crucial for soil management and sustainable agriculture. So far, many attempts have been made to map soil carbon pools with environmental variables through the digital soil mapping (DSM) procedure; nevertheless, little attempt has been made to integrate remote sensing and proximal sensing for this purpose. The main goal of this study was to predict carbon pools by the integration of numerous auxiliary variables (Terrain variables and remotely sensed data) with Vis-NIR-SWIR spectral data and magnetic measures at the semiarid region in the west of Iran at the watershed scale. A total of 346 samples were collected in a randomly manner from the surface soil ( 0–20 cm). Visible-near-infrared (Vis-NIR) data derived in the ranges of 350–2500 nm, magnetic susceptibility (χ_hf, χ_lf, and χ_fd), and six soil properties (CCE, EC, pH, Sand, Silt and clay) comprising carbon pools were measured. Also, a total of thirty-two environmental variables were derived from topographic features and remote sensing data. For DSM modeling and preparation of a digital map of the carbon pools, two models (random forest (RF) and Cubist) were considered in three scenarios (i) remote sensing indices + topographic attributes + soil properties; ii) variables in the first scenario + magnetic measures; iii) variables in second scenario + spectral data). The result showed that integrating proximal sensing and remote sensing (scenario III) provided better performance in both models. Also, the results showed that the Cubist model compared to the RF model provided a slightly higher performance in estimating soil organic carbon (SOC) (R²=0.67, RMSE=0.26) and soil inorganic carbon (SIC) (R²=0.66, RMSE=0.76). However, for predicting the total soil carbon (TSC) content, the RF model with a coefficient of determination of 0.66, and root mean square error (RMSE) of 0.81 performed better than the Cubist model (R²=0.60, RMSE=0.85). The furthermost imperative variables in the spatial estimation of soil carbon stocks were proximal attributes, remote sensing attributes, soil properties, and digital elevation model (DEM) derivatives, respectively. Overall, the results demonstrated the high capability of the integration of proximal soil sensing and remote sensing to predict soil carbon stocks on a large scale.