Utilizing Image Segmentation with Supervised Machine Learning to Analyze Air Particulate Samples
Faculty Mentor Information
Dr. Sandra Miller
Presentation Date
7-2017
Abstract
Air quality samples are taken at hundreds of locations around the world on a daily basis. These samples are typically quantified using gravimetric methods and classified using chemical techniques. The purpose of this project was to establish a systematic algorithmic procedure using digital image characterization to quantify particle samples with greater accuracy and precision than established techniques. A model with a user-controlled interface was created that has an automated processing system component that does not require any user-supplied data after training is performed. To test the performance of the model, samples of particles ranging from 18 µm to 0.056 µm in diameter were collected from the roof of the BSU Environmental Research Building every 24 hours over a period of 55 days. Two Microorfice Uniform Deposit Impactors (MOUDI) were used, which allowed particulates to deposit in a fashion such that groupings of particles smaller than 18 µm were visible on aluminum substrates without visual enhancement. Programming results showed this is an accurate method for determining surface area coverage for a variety of particle deposit patterns, but is insufficient for quantifying total particle volume or mass.
Utilizing Image Segmentation with Supervised Machine Learning to Analyze Air Particulate Samples
Air quality samples are taken at hundreds of locations around the world on a daily basis. These samples are typically quantified using gravimetric methods and classified using chemical techniques. The purpose of this project was to establish a systematic algorithmic procedure using digital image characterization to quantify particle samples with greater accuracy and precision than established techniques. A model with a user-controlled interface was created that has an automated processing system component that does not require any user-supplied data after training is performed. To test the performance of the model, samples of particles ranging from 18 µm to 0.056 µm in diameter were collected from the roof of the BSU Environmental Research Building every 24 hours over a period of 55 days. Two Microorfice Uniform Deposit Impactors (MOUDI) were used, which allowed particulates to deposit in a fashion such that groupings of particles smaller than 18 µm were visible on aluminum substrates without visual enhancement. Programming results showed this is an accurate method for determining surface area coverage for a variety of particle deposit patterns, but is insufficient for quantifying total particle volume or mass.