Multiscale and multisource data fusion for full-coverage PM_2.5 concentration mapping: Can spatial pattern recognition come with modeling accuracy?

In spite of a variety of PM_2.5 modeling schemes, general guidance for full-coverage PM_2.5 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 PM_2.5 distribution patterns without compromising modeling accuracy. In this study, ten different PM_2.5 concentration data sets were generated using distinct mapping strategies and compared to one another, aiming to facilitate full-coverage PM_2.5 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 PM_2.5 autocorrelation terms as predictors can markedly improve the modeling accuracy, spatial patterns of PM_2.5 estimations could be apparently distorted under different parameter configurations. In an attempt to balance the conflicting objectives, the optimal PM_2.5 mapping scheme was proposed for broadened applications. A daily full-coverage PM_2.5 concentration data set with 5-km resolution in China between 2015 and 2020 was generated for a demonstration to infer an apparent decreasing trend of PM_2.5 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 PM_2.5 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 framework in support of various full-coverage PM_2.5 mapping practices.