PM2.5 concentration simulation by hybrid machine learning based on image features

Air pollution significantly impacts human health, making the development of effective pollutant concentration assessment methods crucial. This study introduces a hybrid machine learning approach to simulate PM2.5 mass concentration using outdoor images, offering an alternative to traditional observation techniques. The proposed method utilizes a convolutional neural network (CNN) to extract image features through transfer learning. The importance of these features is then evaluated using a random forest (RF) model. In addition, the extracted image features are combined with meteorological data (e.g., temperature (TEM), relative humidity (RHU), and sea level pressure (PRS_Sea)) and pollutant concentration data (hourly PM2.5 concentrations from four monitoring stations) for complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) signal decomposition. This results in multiscale signals that are subsequently used in the hybrid machine learning model to simulate PM2.5 concentrations. The results demonstrate that the ResNet50 training method, which extracts 64 image features, yields the best performance. An RF model is applied to the low-frequency signal, superimposed with the trend signal, while a Lasso regression model is used for the high-frequency signal. The combined approach produces superior simulation results than the RF model alone. Notably, image feature 23, PM2.5 concentration from the Institute of Biological Products (IBP), and TEM are most influential for the high-frequency signal, with characteristic coefficients of 1.409, 0.380, and 0.318, respectively. For the low-frequency signals, image features 5 and 23, along with the PM2.5 concentration from the Lanlian Hotel (LH), are the most significant, with importance values of 0.170, 0.137, and 0.125, respectively. The Lasso regression model (random forest model) has the function of high (low) value correction for high (low) frequency signal simulation, leading to more accurate simulation results. This study proposes a cost-effective method for accurately estimating PM2.5 concentrations.