Multiscale and multisource data fusion for full-coverage PM2.5 concentration mapping: Can spatial pattern recognition come with modeling accuracy?

In spite of a variety of PM2.5 modeling schemes, general guidance for full-coverage PM2.5 concentration mapping from satellite observations is still lacking. The current technical gap is tied to how to integrate multiscale data from multiple sources to generate a spatially contiguous map that can better recognize PM2.5 distribution pat-terns without compromising modeling accuracy. In this study, ten different PM2.5 concentration data sets were generated using distinct mapping strategies and compared to one another, aiming to facilitate full-coverage PM2.5 concentration mapping with a generalized approach. The inter-comparison results indicated that different mapping strategies could yield comparable modeling accuracy albeit distinct PM(2.5)distributions over space. Although the inclusion of PM2.5 autocorrelation terms as predictors can markedly improve the modeling accuracy, spatial patterns of PM2.5 estimations could be apparently distorted under different parameter configurations. In an attempt to balance the conflicting objectives, the optimal PM2.5 mapping scheme was proposed for broadened applications. A daily full-coverage PM2.5 concentration data set with 5-km resolution in China be-tween 2015 and 2020 was generated for a demonstration to infer an apparent decreasing trend of PM2.5 across China over the past five years. Besides, the examination of COVID-19 pandemic impacts on regional air quality variations reveals a pattern of marked PM2.5 concentration decrease that cannot be easily realized by site-based air quality measurements. It is indicative that the proposed approach in this study can offer an optimal frame-work in support of various full-coverage PM2.5 mapping practices.