Machine learning method using camera image patterns for predictions of particulate matter concentrations

Abstract

A machine learning technique using image patterns has been proposed to predict the concentrations of particulate matter with an aerodynamic diameter of <10 μm (PM10) originating from on-road vehicles. The analysis process included measurements of PM10 concentrations, image data collection on vehicle transportation patterns and convolutional neural network (CNN). Preprocessing procedures involving high-pass filtering were used to remove the outliers from the measured concentrations of PM10, and multi-neuron components were modeled by connecting the estimated neuron functions. Finally, PM10 concentrations generated from vehicles were predicted through decisions from neural network modeling made in correspondence with the imagery data on estimated vehicles. Machine learning was performed by the classification of four types of images and the results showed good agreement with measurements for the cases of usage in the image patterns of vehicles. Machine learning applied to images of camera photography in this study was designed specifically for the prediction of PM10 concentrations. The findings showed that the proposed method can be used to estimate the real-time contributions of PM10 by vehicles near a road through connection with traffic information from a closed circuit television (CCTV).