Highly Efficient Picture-Based Prediction of PM2.5 Concentration

Air pollutants do much harm to human safety. In particular, the fine particulate matter (PM2.5), a complex air pollutant which is composed of the particles beneath the aerodynamic diameters of 2.5 μm, very possibly causes severe diseases since it is easy to intrude into the lungs. To that end, in this paper, we design a new picture-based predictor of PM2.5 concentration (PPPC), which employs the pictures acquired using mobile phones or cameras to make a real-time estimation of PM2.5 concentration. First, using a large body of pictures which were captured under the good weather conditions, i.e., very low PM2.5 concentration, naturalness statistics (NS) models are built upon entropy features in spatial and transform domains. Second, for a novel picture, we measure its deviation degree from the above-mentioned NS models, considering the fact that the naturalness of a picture tends to reduce with the PM2.5 concentration increased. Third, a simple nonlinear function is introduced to map the deviation degree to the PM2.5 concentration. In comparison to existing relevant state-of-theart predictors, sufficient experimental results manifest the superiority of the proposed PPPC model in terms of prediction accuracy and implementation efficiency.