Spatiotemporal air quality prediction using stochastic advection–diffusion model for multimodal data fusion

Abstract Particulate matter poses significant risks to respiratory and cardiovascular health. Monitoring ambient particulate matter concentrations can provide information on potential exposures and inform mitigation strategies, but ground-based measurements are sparse. Data fusion approaches that integrate data from multiple sources can complement existing observation networks and reveal insights that single-sensor data might miss to better manage pollutant exposure risks. However, data fusion approaches face multiple challenges, including incompatible measurement units, varying data resolutions, and differing levels of uncertainty. As a result, the optimal method for data fusion remains an open question. Here, we propose a probabilistic spatiotemporal model, based on the stochastic advection–diffusion (SAD) equation, as a data fusion method to process multimodal air quality data to predict hourly concentrations of fine particulate matter (PM 2.5 ). We employ a variational inference method to calibrate the probabilistic model using ground-level observations and the numerical output of two simulation models. We then evaluate the prediction performance of our model for two scenarios: (1) incorporating simulation outputs and ground-level observations from sparse regulatory-grade stations and (2) using ground-level observations from both low-cost and regulatory-grade stations. For the first scenario, the data fusion method reduces prediction error by 14% compared to the nearest regulatory-grade air monitor located 20 km away. For the second scenario, error is reduced by 40% compared to the nearest regulatory-grade monitor and 11% compared to the nearest low-cost sensor located approximately 1 km away. The model captures 78% of observed data within a 75% confidence interval across both scenarios, demonstrating its ability to accurately represent uncertainty. Our findings demonstrate that the proposed SAD model can effectively integrate multimodal data to provide improved prediction of particulate matter concentrations at high spatial resolution. Model outputs can inform individual and community-level decision-making to mitigate air pollutant exposures.