Prediction of Soil Organic Carbon at Field Scale by Regression Kriging and Multivariate Adaptive Regression Splines Using Geophysical Covariates

Daniela De Benedetto,Emanuele Barca,Giovanni Lacolla,Stefano Popolizio,Anna Maria Stellacci,Mirko Castellini

doi:10.3390/land11030381

Abstract

Knowledge of the spatial distribution of soil organic carbon (SOC) is of crucial importance for improving crop productivity and assessing the effect of agronomic management strategies on crop response and soil quality. Incorporating secondary variables correlated to SOC allows using information often available at finer spatial resolution, such as proximal and remote sensing data, and improving prediction accuracy. In this study, two nonstationary interpolation methods were used to predict SOC, namely, regression kriging (RK) and multivariate adaptive regression splines (MARS), using as secondary variables electromagnetic induction (EMI) and ground-penetrating radar (GPR) data. Two GPR covariates, representing two soil layers at different depths, and X geographical coordinates were selected by both methods with similar variable importance. Unlike the linear model of RK, the MARS model also selected one EMI covariate. This result can be attributed to the intrinsic capability of MARS to intercept the interactions among variables and highlight nonlinear features underlying the data. The results indicated a larger contribution of GPR than of EMI data due to the different resolution of EMI from that of GPR. Thus, MARS coupled with geophysical data is recommended for prediction of SOC, pointing out the need to improve soil management to guarantee agricultural land sustainability.

Highlights

Soil organic carbon (SOC) is one of the most important indicators for assessing soil quality and overall soil health [1]
Descriptive statistics showed that SOC data were normally distributed as confirmed by skewness and kurtosis values (Table 1) and by Shapiro–Wilk test (p = 0.656); for this reason, they were not subjected to a normal transform
The results of our investigation showed that multivariate adaptive regression splines (MARS) outperformed regression kriging (RK) in predicting SOC spatial distribution

Summary

Introduction

Soil organic carbon (SOC) is one of the most important indicators for assessing soil quality and overall soil health [1]. Because of the interaction of the factors described, SOC spatial variation is often wide and complex, and the knowledge of its spatial distribution is the key information in agricultural productivity to improve food security, enhance crop production [11], and predict the effects of different agronomic management strategies. Among these strategies, irrigation with treated municipal wastewater can be considered important for saving limited freshwater resources and protecting the environment, but its effects should be monitored to avoid soil fertility decline in the medium to long term [9]

Methods

Results

Discussion

Conclusion