Air Quality Estimation and Forecasting via Data Fusion With Uncertainty Quantification: Theoretical Framework and Preliminary Results

Integrating air quality information from models, satellites, and in situ monitors allows for both better estimation of air quality and better quantification of uncertainties in this estimation. Uncertainty quantification is important to appropriately convey confidence in these estimates and forecasts to users who will base decisions on these. Uncertainty quantification also allows tracing the value of information provided by different data sources. This can identify gaps in the monitoring network where additional data could further reduce uncertainties. This paper presents a framework for data fusion with uncertainty quantification, applicable to multiple air‐quality‐relevant pollutants. Testing of this framework in the context of nitrogen dioxide forecasting at sub‐city scales shows promising results, with confidence intervals typically encompassing the expected number of actual measurements during cross‐validation. The framework is now being implemented into an online tool to support local air quality management decision‐making. Future work will also include the incorporation of low‐cost air sensor data and the quantification of uncertainty at hyper‐local scales. Plain Language Summary Poor air quality has adverse impacts on human and environmental health. Estimating and forecasting air quality accurately can improve early warnings and mitigation for poor air quality. Furthermore, understanding the uncertainties and degree of confidence in these forecasts and estimates can help air quality managers know when and where they can be relied upon, and where more data might still be needed. This paper outlines a method to combine air quality information from models, satellites, and ground‐based monitors, and to assess the confidence in the combined output. Combining all these data sources can give us a better overall understanding of air quality, and making comparisons between them allows us to better understand uncertainties. We find that the method can produce reasonable assessments of the confidence it has in its estimates, with the expected numbers of actual measurements usually falling within the confidence intervals produced by the method. An exception is when this method is applied very close to a major pollution source (e.g., a highway, in our study). In such cases, since the method does not know that there is such a source nearby, it tends to be overconfident in its prediction. Key Points The proposed data fusion method also produces uncertainty asses