Performance evaluation of photographic measurement in the machine-learning prediction of ground PM2.5 concentration

Abstract

The scarcity and sparsity of meteorological vertical observations have caused severe restrictions in the accurate forward prediction and backward analysis of weather and air pollution. In this study, we used a common camera to automatically photograph the daytime and nighttime lights at an urban site in Beijing during 2019–2020, to support or act as a possible alternative to radiosonde measurement. The photo features characterize the scattering effect of atmospheric particulate matter on visible sunlight or lamps and record cloud, fog, and precipitation processes. The probability distributions of the mean blue brightness (B) and the mean red brightness (R) are significantly different for both the sky-part and the ground-part pixels of the photo. The B/R ratio of the sky part (B/R_sky) is exponentially and negatively correlated with ground PM2.5 concentration. During the daytime, B/R_sky has a higher priority than the boundary layer height (BLH) in the determination of PM2.5 concentration by a decision tree model, whereas the BLH plays a key role at night, and the importance of B/R_sky is comparable to that of the BLH in the decision tree. The photo features were adopted as input variables into a simple multi-layer perceptron (MLP) model (3-layer neural network) predicting PM2.5 concentration, with a root mean square error of 1–3 μg/m3, indicating that the auto-shooting camera is a competitive alternative of meteorological measurement. Because of the low cost of installation, the broad application of auto-photography can make up for deficiencies of sounding observation, i.e., sparse temporal and spatial resolution, and can be accessed by the public in real time.