On the Image Processing Methods for Morphological Characterization of Biomass Smoke Particles

HAMED NIKOOKAR, Timothy Sipkens, Steven Rogak, University of British Columbia

     Abstract Number: 527
     Working Group: Biomass Combustion: Outdoor/Indoor Transport and Indoor Air Quality

Abstract
Image processing methods have found wide use in interpretation of microscopic micro/nanoscale data. For the special case of soot, they are implemented in conjunction with Transmission Electron Microscopy (TEM) to study different morphological properties such as aggregate equivalent size, primary particle size, number of primaries, and fractal properties. Recently, the fast automated image segmentation and primary particle detection algorithms have replaced the time and effort taking user-adjusted analyses in the study of soot. The automated approaches include both the more classic methods that involve filtering, transformation, and morphological operations calibrated for certain applications, and the newer unsupervised and supervised machine learning methods that self-tune the operations discussed. While the previous studies showed that these methods are more or less robust for fresh soot, they have not been explored so much for the scenarios where the emissions age in the ambient air and produce quite diverse particle populations. In this study, we investigated an example of such cases where the brown haze images from a forest fire near Kamloops, BC in July 2021 were examined using a recently developed machine learning method that implements k-means clustering to decide on the segmentation of particles from their backgrounds and prepares them for further higher-level analyses. The segregated particles are evaluated by their size and three other less-studied parameters, i.e., circularity, optical depth, and acutance- that account for the shape and radiative properties. Results based on k-means are compared with the manual ones to assess the performance of this state-of-the-art method for more complex particle populations. It is shown that k-means interestingly performs well enough to capture particle boundaries and well predicts the second-hand morphological distributions. Discrepancies, however, also appear from the manual data especially for the cases where the effect of aging is more dominant.