Ultrahigh-resolution PM estimation from top-of-atmosphere reflectance with machine learning: Theories, methods, and applications.

Intra-urban pollution monitoring requires fine particulate (PM) concentration mapping at ultrahigh-resolution (dozens to hundreds of meters). However, current PM concentration estimation, which is mainly based on aerosol optical depth (AOD) and meteorological data, usually had a low spatial resolution (kilometers) and severe spatial missing problem, cannot be applied to intra-urban pollution monitoring. To solve these problems, top-of-atmosphere reflectance (TOAR), which contains both the information about land and atmosphere and has high resolution and large spatial coverage, may be efficiently used for PM estimation. This study aims to systematically evaluate the feasibility of retrieving ultrahigh-resolution PM concentration at a large scale (national level) from TOAR. Firstly, we make a detailed discussion about several important but unsolved theoretic problems on TOAR-based PM retrieval, including the band selection, scale effect, cloud impact, and mapping quality evaluation. Secondly, four types and eight retrieval models are compared in terms of quantitative accuracy, mapping quality, model generalization, and model efficiency, with the pros and cons of each type summarized. Deep neural network (DNN) model shows the highest retrieval accuracy, and linear models were the best in efficiency and generalization. As a compromise, ensemble learning shows the best overall performance. Thirdly, using the highly accurate DNN model (cross-validated R equals 0.93) and through combining Landsat 8 and Sentinel 2 images, a 90 m and ∼4-day resolution PM product was generated. The retrieved maps were used for analyzing the fine-scale interannual pollution change inner the city and the pollution variations during novel coronavirus disease 2019 (COVID-19). Results of this study proves that ultrahigh resolution can bring new findings of intra-urban pollution change, which cannot be observed at previous coarse resolution. Lastly, some suggestions for future ultrahigh-resolution PM mapping research were given.