The prediction of soil chemical and physical properties from mid-infrared spectroscopy and combined partial least-squares regression and neural networks (PLS-NN) analysis

L. J. Janik, S. T. Forrester, Graeme Rawson

Research output: Contribution to journalArticle

74 Citations (Scopus)

Abstract

This study compares the performance of partial least-squares (PLS) regression analysis for the prediction of a wide range of soil chemical and physical properties from their mid-infrared (MIR) spectra with those from a combination of PLS regression and neural networks (NN). A combination of PLS and NN, referred to as PLSNN, uses a relatively few PLS scores as inputs to the NN. This has the advantages of the robustness and qualitative and quantitative features of PLS and the non-linear capabilities of neural networks. In this study, the PLS-NN method outperformed the basic PLS regression for the prediction of some soil properties from MIR spectra of soils from throughout New South Wales, Australia. The coefficient of determination (R2) and root-mean standard error of prediction (RMSEP) for total organic carbon (TOC) were improved from an R2=0.87 and RMSEP=0.7 by PLS, to an R2=0.94 and RMSEP=0.5 by PLS-NN. Predictions appeared to be most improved where PLS regression curvature was apparent or when the analytical value distributions were heavily skewed to low values with many negative predictions, e.g. for exchangeable-Al and Mg, TOC, CEC, Psorption and moisture content. For other soil properties there appeared to be a significant improvement in the lower 25 percentile of analyte values but the prediction errors increased at high analyte values and no net improvement was achieved, or were only marginal for the full range of data. In cases where PLS failed to give a viable model, NN also could not derive a converging model. The use of PLS-NN over the usual PLS method for routine soil analytical applications must be questioned with regard to a tradeoff between some possible limited improvement versus the added computational complexity and additional software requirement for NN.
Original languageEnglish
Pages (from-to)179-188
Number of pages10
JournalChemometrics and Intelligent Laboratory Systems
Volume97
Issue number2
DOIs
Publication statusPublished - 2009

Fingerprint

Electric network analysis
Chemical properties
Infrared spectroscopy
Physical properties
Neural networks
Soils
Organic carbon
Infrared radiation
Regression analysis
Computational complexity
Moisture

Cite this

@article{66ea04c455514b258351a7d721946b11,
title = "The prediction of soil chemical and physical properties from mid-infrared spectroscopy and combined partial least-squares regression and neural networks (PLS-NN) analysis",
abstract = "This study compares the performance of partial least-squares (PLS) regression analysis for the prediction of a wide range of soil chemical and physical properties from their mid-infrared (MIR) spectra with those from a combination of PLS regression and neural networks (NN). A combination of PLS and NN, referred to as PLSNN, uses a relatively few PLS scores as inputs to the NN. This has the advantages of the robustness and qualitative and quantitative features of PLS and the non-linear capabilities of neural networks. In this study, the PLS-NN method outperformed the basic PLS regression for the prediction of some soil properties from MIR spectra of soils from throughout New South Wales, Australia. The coefficient of determination (R2) and root-mean standard error of prediction (RMSEP) for total organic carbon (TOC) were improved from an R2=0.87 and RMSEP=0.7 by PLS, to an R2=0.94 and RMSEP=0.5 by PLS-NN. Predictions appeared to be most improved where PLS regression curvature was apparent or when the analytical value distributions were heavily skewed to low values with many negative predictions, e.g. for exchangeable-Al and Mg, TOC, CEC, Psorption and moisture content. For other soil properties there appeared to be a significant improvement in the lower 25 percentile of analyte values but the prediction errors increased at high analyte values and no net improvement was achieved, or were only marginal for the full range of data. In cases where PLS failed to give a viable model, NN also could not derive a converging model. The use of PLS-NN over the usual PLS method for routine soil analytical applications must be questioned with regard to a tradeoff between some possible limited improvement versus the added computational complexity and additional software requirement for NN.",
keywords = "Author Keywords: PLS-NN, CALIBRATION, CLASSIFICATION, EXTRACTION, KeyWords Plus: REFLECTANCE SPECTROSCOPY, Mid-infrared, Neural networks, PLS, SPECTRA, Soils",
author = "Janik, {L. J.} and Forrester, {S. T.} and Graeme Rawson",
note = "Imported on 12 Apr 2017 - DigiTool details were: Journal title (773t) = Chemometrics and Intelligent Laboratory Systems. ISSNs: 0169-7439;",
year = "2009",
doi = "10.1016/j.chemolab.2009.04.005",
language = "English",
volume = "97",
pages = "179--188",
journal = "Chemometrics and Intelligent Laboratory Systems",
issn = "0169-7439",
publisher = "Elsevier",
number = "2",

}

TY - JOUR

T1 - The prediction of soil chemical and physical properties from mid-infrared spectroscopy and combined partial least-squares regression and neural networks (PLS-NN) analysis

AU - Janik, L. J.

AU - Forrester, S. T.

AU - Rawson, Graeme

N1 - Imported on 12 Apr 2017 - DigiTool details were: Journal title (773t) = Chemometrics and Intelligent Laboratory Systems. ISSNs: 0169-7439;

PY - 2009

Y1 - 2009

N2 - This study compares the performance of partial least-squares (PLS) regression analysis for the prediction of a wide range of soil chemical and physical properties from their mid-infrared (MIR) spectra with those from a combination of PLS regression and neural networks (NN). A combination of PLS and NN, referred to as PLSNN, uses a relatively few PLS scores as inputs to the NN. This has the advantages of the robustness and qualitative and quantitative features of PLS and the non-linear capabilities of neural networks. In this study, the PLS-NN method outperformed the basic PLS regression for the prediction of some soil properties from MIR spectra of soils from throughout New South Wales, Australia. The coefficient of determination (R2) and root-mean standard error of prediction (RMSEP) for total organic carbon (TOC) were improved from an R2=0.87 and RMSEP=0.7 by PLS, to an R2=0.94 and RMSEP=0.5 by PLS-NN. Predictions appeared to be most improved where PLS regression curvature was apparent or when the analytical value distributions were heavily skewed to low values with many negative predictions, e.g. for exchangeable-Al and Mg, TOC, CEC, Psorption and moisture content. For other soil properties there appeared to be a significant improvement in the lower 25 percentile of analyte values but the prediction errors increased at high analyte values and no net improvement was achieved, or were only marginal for the full range of data. In cases where PLS failed to give a viable model, NN also could not derive a converging model. The use of PLS-NN over the usual PLS method for routine soil analytical applications must be questioned with regard to a tradeoff between some possible limited improvement versus the added computational complexity and additional software requirement for NN.

AB - This study compares the performance of partial least-squares (PLS) regression analysis for the prediction of a wide range of soil chemical and physical properties from their mid-infrared (MIR) spectra with those from a combination of PLS regression and neural networks (NN). A combination of PLS and NN, referred to as PLSNN, uses a relatively few PLS scores as inputs to the NN. This has the advantages of the robustness and qualitative and quantitative features of PLS and the non-linear capabilities of neural networks. In this study, the PLS-NN method outperformed the basic PLS regression for the prediction of some soil properties from MIR spectra of soils from throughout New South Wales, Australia. The coefficient of determination (R2) and root-mean standard error of prediction (RMSEP) for total organic carbon (TOC) were improved from an R2=0.87 and RMSEP=0.7 by PLS, to an R2=0.94 and RMSEP=0.5 by PLS-NN. Predictions appeared to be most improved where PLS regression curvature was apparent or when the analytical value distributions were heavily skewed to low values with many negative predictions, e.g. for exchangeable-Al and Mg, TOC, CEC, Psorption and moisture content. For other soil properties there appeared to be a significant improvement in the lower 25 percentile of analyte values but the prediction errors increased at high analyte values and no net improvement was achieved, or were only marginal for the full range of data. In cases where PLS failed to give a viable model, NN also could not derive a converging model. The use of PLS-NN over the usual PLS method for routine soil analytical applications must be questioned with regard to a tradeoff between some possible limited improvement versus the added computational complexity and additional software requirement for NN.

KW - Author Keywords: PLS-NN

KW - CALIBRATION

KW - CLASSIFICATION

KW - EXTRACTION

KW - KeyWords Plus: REFLECTANCE SPECTROSCOPY

KW - Mid-infrared

KW - Neural networks

KW - PLS

KW - SPECTRA

KW - Soils

U2 - 10.1016/j.chemolab.2009.04.005

DO - 10.1016/j.chemolab.2009.04.005

M3 - Article

VL - 97

SP - 179

EP - 188

JO - Chemometrics and Intelligent Laboratory Systems

JF - Chemometrics and Intelligent Laboratory Systems

SN - 0169-7439

IS - 2

ER -