A CatBoost approach with wavelet decomposition to improve satellite-derived high-resolution PM2.5 estimates in Beijing-Tianjin-Hebei. (15th March 2021)

• Record Type: Journal Article
• Authors: Ding, Yu; Chen, Zuoqi; Lu, Wenfang; Wang, Xiaoqin
• Is Part Of: Atmospheric environment. Volume 249(2021)
• Publication Date: 2021-03-15
• Subjects: PM2.5 -- AOD -- CatBoost -- Wavelet decomposition -- Remote sensing -- Beijing-Tianjin-Hebei; Air -- Pollution -- Periodicals; Air -- Pollution -- Meteorological aspects -- Periodicals; 551.51
• DOI: 10.1016/j.atmosenv.2021.118212
• Languages: English
• ISSNs: 1352-2310

Abstract

Abstract: High-resolution data of fine particulate matters (PM2.5 ) are of great interest for air pollution prevention and control. However, due to the uneven spatial distribution of ground stations, satellite acquisition cycle, and cloud/rain, high-resolution products cannot be provided on a complete spatio-temporal scale. To provide a full daily PM2.5 product in recent years at 1-km grid of the Beijing-Tianjin-Hebei (BTH) region, here we apply a state-of-the-art machine learning approach, CatBoost, to (1) reconstruct satellite aerosol optical depth (AOD) data; and to (2) estimate gridded PM2.5 from station measurements combining elevation, meteorological factors, and the reconstructed AOD data. Compared with existing approaches, CatBoost substantially improved the performance of AOD reconstruction by ~16%. We further show that a wavelet decomposition procedure on the station-based PM2.5 and input variables is helpful to improve the estimation accuracy. Overall, the approach has a good performance in estimating PM2.5 with a cross-validation R 2 of 0.88 and root-mean-squared error of 17.79 μg/m 3 . From the new dataset, population-weighted PM2.5 revealed heterogeneous spatial distribution of exposure in different areas, consistently higher in the Southern and Eastern BTH and lower in Beijing and Northern BTH. In recent years, both AOD and PM2.5 in the BTH region had notable interannual decreases, which can be attributed to the emission reduction efforts and interannual natural variabilities. Highlights: A new method was applied to estimate fine particulate matters from remote sensing. The new method was based on a novel machine learning method, CatBoost. The optimized method included a wavelet decomposition procedure. Our method outperformed existing methods compared with station observations.