Automatic Image-based Air Quality Detection based on Deep Convolutional Neural Networks and Histogram Analysis

By growing industrial cities, air pollution has become one of the most important issues in the world. Since air quality impacts on human health directly, creating automatic systems to determine air pollution is so important. Frequent episodes of unhealthy air pollution conditions have been reported for Tehran, the capital of Iran, mainly because of critically high levels of fine particulate matter (PM2.5). This study aims to estimate fine particulate matter (PM2.5) concentration using RGB images processing in Tehran, Iran. The proposed method uses hybrid models of deep convolutional neural networks and comprehensive histogram of pixels to extract deep and robust features from color images. To train and test the proposed model, we create a new database by combining HVAQ database (which is the first publicly online dataset and has high-resolution images for air quality detection) and some arbitrary images of the sky which have been captured from Tehran city, in different pollution situations. We collected this image from news websites and usual blogs. Based on the date when each picture is taken, we find the relevant PM2.5 factors from the weather site. Six different days were selected belonging to three different quality categories, Good, Moderate and Unhealthy for Sensitive Groups. By combining these images with the HVAQ database, we create a set with three usual categories. By adding these images, the database has more variety and also the number of samples in each class got increased which helps model to be trained better. Results demonstrate that the proposed method performs better than the state-of-the art and can automatically detect the PM2.5 concentrations from color images.