Premium
Determination of Soil Organic Matter and Carbon Fractions in Forest Top Soils using Spectral Data Acquired from Visible–Near Infrared Hyperspectral Images
Author(s) -
O'Rourke S. M.,
Holden N. M.
Publication year - 2012
Publication title -
soil science society of america journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.836
H-Index - 168
eISSN - 1435-0661
pISSN - 0361-5995
DOI - 10.2136/sssaj2011.0053
Subject(s) - hyperspectral imaging , soil water , environmental science , soil carbon , total organic carbon , soil organic matter , soil test , organic matter , soil science , environmental chemistry , remote sensing , chemistry , geology , organic chemistry
Adequately quantifying C sequestration in soil, post 2012, can be used to offset C losses in national greenhouse gas inventory but requires very large sample numbers and rapid analytical methods. Wet and dry combustion methods are analytically accurate but expensive and slow while optical techniques have the potential to provide rapid, cost‐effective alternatives. This study examined the potential of spectral data acquired from laboratory hyperspectral imaging (HSI) systems and chemometric analysis to predict soil organic matter (OM), total carbon (TC), inorganic carbon (IC), and organic carbon (OC) fractions in forest top soils from Avondale Forest Park, Rathdrum, County Wicklow, Ireland. The spectral range of hyperspectral instruments operating in the visible (VIS; 400–1000 nm), near infrared (NIR; 880–1720 nm) and combined VIS–NIR regions (400–1720 nm) were investigated for each soil property. Validations using a randomly selected 25% partition of the dataset indicated that the best soil TC and OC predictions were achieved in the VIS region, a ratio of predicted deviation (RPD) indicated excellent predictions for both TC (3.39) and OC (3.39). The best OM and IC prediction was achieved in the VIS–NIR region, OM ranked as excellent (3.06) but IC produced models with very poor predictive ability (1.26) due to a limited range of concentrations. Model robustness was tested using alternative methods of partitioning the dataset ( n = 152). Partitioning following stratification by TC or OC concentration improved accuracy by 1.4‐fold, while soil OM accuracy was improved 1.2‐fold after stratifying by sampling site. When independent validations were tested on “new sites” by holding each sampling site out of model calibration in turn, OC predicted with reasonable root mean square error (RMSE) for most sites but produced RPD values indicating poor predictive performance. A certain degree of uniqueness associated with soils at new sites caused model accuracy to deteriorate. Overall results indicate that there is much potential to develop hyperspectral imaging as a methodology for soil C and OM analyses, but soils from the intended target site must be included in the model calibration to maintain model prediction accuracy.