Deep Learning-Based PM2.5 Concentration Estimation Using Convolutional Neural Network on Environmental Image Data

Globalization and rapid progress in technology is leading to the pollution of the environment severely affecting public health and leading to deaths. Air quality in most of the cities at global level are is very poor and countries are taking several measures to ensure that clean air is available to the public. One pollutant that is to be taken care of in this regard is PM2.5 (Particulate Matter 2.5). In general, estimation of PM2.5 from the environment is measured by using sensors deployed at various locations. Atmospheric pollution also affects the quality of the images captured there. In this work, a novel algorithm using deep learning technique is proposed for estimating the PM2.5 level in the environment by using images captured by the cameras. This problem is forecasted as a regression problem. The features are extracted from the selected pre-trained networks and the corresponding PM2.5 values are given to a regressor. The datasets have images and their PM2.5 values collected from three prominent places (Shadipur, ITO, and DTU) in New Delhi, India. Images were captured by webcams and the corresponding PM2.5 values are obtained from the Pollution Control Board. Coefficient of Determination $R^{2}$ is computed as the performance metric. An $R^{2}$ value in excess of 0.91 is obtained on all the three datasets. PM2.5 distribution is not uniform, instead it is skewed. Hence, oversampling techniques are used to overcome the under represented samples.