Image-based PM10 Concentration Classification with Convolutional Neural Network

Abstract

Air pollution is a crucial environmental problem. One of the air quality indicators is the concentration of PM10, which is particulate matter less than 10 microns in size. Unfortunately, PM10 monitoring in Indonesia is only available at 15 observation stations. In fact, the availability of PM10 information in real time is important for pollution control and health protection. Therefore, this research focuses on sky image-based PM10 concentration analysis by utilizing photos taken using smartphone cameras. By applying CNN, PM10 concentration estimation can be done in real time. To achieve the goal, a total of 300 image data were retrieved from Beijing tourism web along with PM10 attributes. The images were classified into three categories: 'good', 'moderate', and 'unhealthy'. A total of 80% of the data is used to train six variations of CNN models. Furthermore, the model with the highest accuracy will be selected as the best model. The results show that CNN architecture with Leaky ReLU activation function and average pooling is valid to classify images based on PM10 concentration. The results of this study can be a powerful tools for improving public health and reducing the impact of air pollution.