PM2.5 Concentration Prediction: Ultrahigh Spatiotemporal Resolution Achieved by Combining Machine Learning and Low-Cost Sensors

PM2.5 pollution is still serious in densely populated cities with frequent traffic activities, and it continues to threaten public health. Therefore, it is urgent that we obtain ultrahigh-resolution data that can reveal its complex spatiotemporal variation characteristics, supporting more refined environmental governance and health risk prevention and control. This study first carried out ground monitoring based on low-cost sensors combined with observation results, which were corrected with the national environmental monitoring station data. This study also introduced multi-source auxiliary variables and constructed a machine learning model through the stacking ensemble learning method. The model combines corrected low-cost sensor data with high-resolution prediction factors to achieve ultrahigh-spatiotemporal-resolution prediction of PM2.5 at 100 m × 100 m spatial resolution and hourly temporal resolution. The results show that the constructed model shows good prediction ability in 5-fold cross validation, with an overall R2 of 0.93 and a root mean square error (RMSE) of 3.09 μg/m3. The spatiotemporal analysis based on the prediction results further revealed that the PM2.5 concentration in the city showed significant variation characteristics at both the ultra-local scale and the short-term scale, reflecting the high heterogeneity of urban air pollution. In addition, by comparing and analyzing the monitoring data of a national environmental monitoring station that were not used in the correction, it was found that the corrected low-cost sensor data significantly reduced the prediction uncertainty, reducing the RMSE from 72.068 μg/m3 to 16.759 μg/m3, verifying its effectiveness in high spatiotemporal resolution air quality monitoring. This shows that low-cost sensors are expected to make up for the problem of insufficient spatial coverage in traditional national environmental monitoring stations, supporting the successful assessment of urban-level air pollution and health risk management, and therefore having broad application prospects.