A model based two-stage classifier for airborne particles analyzed with Computer Controlled Scanning Electron Microscopy

Meier, Mario Federico; Mildenberger, Thoralf; Locher, René; Rausch, Juanita; Zünd, Thomas; Neururer, Christoph; Ruckstuhl, Andreas; Grobéty, Bernard

doi:10.1016/j.jaerosci.2018.05.012

Back

Journal article

+ 2 other files

A model based two-stage classifier for airborne particles analyzed with Computer Controlled Scanning Electron Microscopy

Meier, Mario Federico Department of Geosciences, University of Fribourg, Switzerland - Particle Vision GmbH, c/o Fri Up, Fribourg, Switzerland
Mildenberger, Thoralf Institute for Data Analysis and Process Design (IDP), Zürich University of Applied Sciences Winterthur, Switzerland
Locher, René Institute for Data Analysis and Process Design (IDP), Zürich University of Applied Sciences Winterthur, Switzerland
Rausch, Juanita Department of Geosciences, University of Fribourg, Switzerland - Particle Vision GmbH, c/o Fri Up, Fribourg, Switzerland
Zünd, Thomas Particle Vision GmbH, c/o Fri Up, Fribourg, Switzerland
Neururer, Christoph Department of Geosciences, University of Fribourg, Switzerland
Ruckstuhl, Andreas Institute for Data Analysis and Process Design (IDP), Zürich University of Applied Sciences Winterthur, Switzerland
Grobéty, Bernard Department of Geosciences, University of Fribourg, Switzerland

01.09.2018

Published in:

Journal of Aerosol Science. - 2018, vol. 123, p. 1–16

English Computer controlled scanning electron microscopy (CCSEM) is a widely-used method for single airborne particle analysis. It produces extensive chemical and morphological data sets, whose processing and interpretation can be very time consuming. We propose an automated two-stage particle classification procedure based on elemental compositions of individual particles. A rule-based classifier is applied in the first stage to form the main classes consisting of particles containing the same elements. Only elements with concentrations above a threshold of 5 wt% are considered. In the second stage, data of each main class are isometrically log-ratio transformed and then clustered into subclasses, using a robust, model-based method. Single particles which are too far away from any more densely populated region are excluded during training, preventing these particles from distorting the definition of the sufficiently populated subclasses. The classifier was trained with over 55,000 single particles from 83 samples of manifold environments, resulting in 227 main classes and 465 subclasses in total. All these classes are checked manually by inspecting the ternary plot matrix of each main class. Regardless of the size of training data, some particles might belong to still undefined classes. Therefore, a classifier was chosen which can declare particles as unknown when they are too far away from all classes defined during training.

Faculty

Faculté des sciences et de médecine

Department

Département de Géosciences

Language