High-resolution Spatiotemporal Prediction of PM2.5 Concentration based on Mobile Monitoring and Deep Learning

Abstract

Obtaining the high-resolution distribution characteristics of urban air pollutants is crucial for effective pollution control and public health. In order to fulfill it, mobile monitoring offers a novel and practical approach compared to traditional fixed monitoring methods. However, the sparsity of mobile monitoring data still makes it a challenge to recover the high-resolution pollutant concentration across an entire area. To tackle the sparsity issue and fulfill a prediction of the spatiotemporal distribution of PM_2.5, a high-resolution urban PM_2.5 prediction method was proposed based on mobile monitoring data in this study. This method enables prediction with a spatial resolution of 500m × 500m and a temporal resolution of 1 hour. First, a Light Gradient Boosting Machine (LightGBM) was trained using mobile monitoring of PM_2.5 concentration and exogenous features to obtain complete spatiotemporal PM_2.5 concentration. Second, a model consisting of Convolutional Neural Network and Transformer (CNN-Transformer) with a customised loss function was established to predict high-resolution PM_2.5 concentration based on complete spatiotemporal data. The method was validated using real-world data collected from Cangzhou, China. The numerical results from cross-validation showed an R² of 0.925 for imputation and 0.887 for prediction, demonstrating this method is suitable for high-resolution spatiotemporal prediction of PM_2.5 concentration based on mobile monitoring data.