Improving real-time high-resolution estimates of PM2.5 concentration fields in urban areas by the SmartAQ+ system with data fusion and machine learning

Monitoring PM 2.5 (mass of particles with diameter less than 2.5 μm) concentrations is challenging due to the limited number of ground-level monitoring stations and the limitations of existing modeling approaches. This study introduces SmartAQ+, a data fusion model that combines a chemical transport model-based system (SmartAQ) with low-cost sensor measurements and machine learning (ML) to enhance high-resolution PM 2.5 estimations at the present-time at a 1x1 km 2 scale. SmartAQ+ integrates real-time data from low-cost PM 2.5 sensors, weather stations, and land-use information to improve the accuracy of present-time PM 2.5 estimations at all locations in an urban area. SmartAQ+ demonstrated superior performance compared to SmartAQ that does not use real-time measurements in estimating the present-time PM 2.5 , reducing the corresponding mean error, fractional bias (FBIAS) and fractional error (FERROR) by a factor of two. SmartAQ+ correctly identified 132 out of 190 PM 2.5 exceedance events of the daily limit of 25 μg m -3 , compared to SmartAQ's 34, while reducing false positives by a factor of 2 and missed events by a factor of 3. The performance gains depended on the availability of nearby sensors. In data sparse zones and during unusual events the model can inherit biases from the chemical transport model and can underestimate extremes. The study highlights the potential of data fusion models to address the limitations of standalone approaches, offering more precise air quality estimations in areas of a city in which there are no measurements. • SmartAQ+ improves SmartAQ’s estimations of PM 2.5 with machine learning • It reduced mean error, bias, and fractional error by ∼50% versus SmartAQ baseline • SmartAQ+ detected 70% of PM 2.5 daily limit exceedances, tripling SmartAQ's accuracy • SmartAQ+ performs well in locations without sensor data