Ultrahigh Dimensional Variable Selection for Mapping Soil Carbon

TL;DR

This paper demonstrates the effectiveness of LASSO-based variable selection in ultrahigh dimensional soil carbon mapping, enabling efficient covariate selection from hundreds of variables for spatial interpolation.

Contribution

It introduces the application of LASSO variable selection for ultrahigh dimensional soil mapping, showing its computational efficiency and effectiveness in covariate selection.

Findings

LASSO efficiently selects covariates from 800 variables.

LASSO reduces computational demand in high-dimensional scenarios.

Improves spatial interpolation accuracy for soil carbon.

Abstract

Modern soil mapping is characterised by the need to interpolate samples of geostatistical response observations and the availability of relatively large numbers of environmental characteristics for consideration as covariates to aid this interpolation. We demonstrate the efficiency of the Least Angle Regression algorithm for Least Absolute Shrinkage and Selection Operator (LASSO) penalized multiple linear regression at selecting covariates to aid the spatial interpolation of geostatistical soil carbon observations under an ultrahigh dimensional scenario. Where an exhaustive search of the models that could be constructed from 800 potential covariate terms and 60 observations would be prohibitively demanding, LASSO variable selection is accomplished with trivial computational investment.