0 avis
Soil classification of multi-horizontal profiles using support vector machines and vis-NIR spectroscopy
Archive ouverte
Edité par CCSD -
The need for rapid and inexpensive techniques for high-resolution soil information has led toimprovements over traditional methods, and in particular those based on visible near-infrared(vis–NIR) spectroscopy. While vis–NIR has been used for soil classification for some preliminarystudies, how to combine spectral information from soil profiles remains a substantial challenge.This study was undertaken to investigate the potential of vis–NIR to discriminate soil classes onprofiles containing various soil horizons.We took 130 soil profiles at Zhejiang province, of which were classified in the field at suborderlevel according to Chinese Soil Taxonomy (5 soil orders and 10 suborders). Subsoil samples weretaken by diagnosis layers (A, B and C). Support vector machine (SVM) algorithm was used todetermine the soil classes, by analyzing quantitatively their diffuse reflectance spectra in the vis–NIR range. For SVM is a binary classification algorithm, the qualitative analysis was conducted bycombining the votes of each sample from the same profile and the class got most votes in one profilewas defined as their predicted soil class. Readily available variables (soil color) and well-predictedproperties (soil organic matter, soil texture and pH) using vis-NIR spectra were added as auxiliaryinformation.Using synthesized model (spectra plus auxiliary soil information), SVM produced better clas-sification performances at soil order level and suborder level (accuracy were 68.29% and 63.51%respectively) than spectra independently (accuracy were 60.69% and 58.54% respectively). Theysuggest that vis–NIR spectroscopy combining votes gained from SVM could make an essentialcontribution to the identification of soil classes in an effective approach of soil classification evenwhen profiles contain various soil horizons.
Consulter en ligne
Suggestions
Du même auteur
