| Title | PM2.5 concentration estimation with 1-km resolution at high coverage over urban agglomerations in China using the BPNN-KED approach and potential application |
| Authors | 朱忠敏 |
| Issue Date | 2021 |
| Publisher | Atmospheric Research |
| Keywords | 1-km resolution Back-propagation neural network High-coverage Kriging with external drifting PM2.5 Urban agglomerations |
| Citation | Atmospheric Research. 2021;258:N.PAG. |
| Abstract | Fine particle matter (PM 2.5) significantly affects the atmospheric environment and human health. The satellite-derived aerosol optical depth (AOD), which could represent the concentration of atmospheric particles to a certain extent, is widely used for estimating ambient PM 2.5 concentration, in combination with diverse auxiliary information. However, the general satellite-derived PM 2.5 products exhibit limitation in the application and aggregate analysis of PM 2.5 in urban areas, because of the moderate spatial resolution to match the urban landscape and low spatial coverage making it hard to describe airmass trajectory. In order to explore the potential application value of PM 2.5 concentration products with relatively high spatial coverage and resolution, a two-stage machine learning and geo-statistics coupled model incorporating with a feedback mechanism was proposed in this study. To be specific, we firstly develop a hybrid back-propagation neural network coupled kriging with external drifting approach (BPNN-KED) for estimating 1-km daily PM 2.5 concentration maps at high coverage over four urban agglomerations in China. The model performs well, with R2 up to 0.83 and root mean square error of 14.7 μg/m3 from cross-validation. The daily PM 2.5 maps display an average spatial coverage exceeding 95%, and on an average, each grid produces 350 days of valid estimations annually. In addition, the extra value of the high-coverage PM 2.5 estimates were explored through the more accurate aggregate analysis of urban PM 2.5 pollution level. The advantage of the high-coverage PM 2.5 estimation is demonstrated through daily PM 2.5 hotspot identification over urban areas, providing substantially fine spatially resolved PM 2.5 trends, which offers the potential for daily pollutant emission sources location through satellite remote sensing technology. Moreover, the spatiotemporally continuous PM 2.5 concentrations possess the ability to capture polluted air mass trajectories, thereby offering observational support not only for evaluating the contribution from exogenous pollutants to local PM 2.5 concentrations and but also for providing empirical references for haze warning. • A model coupled machine learning with geostatistics was proposed to estimate PM 2.5 concentrations. • Comprehensive analysis of 1-km PM 2.5 concentration in four urban agglomerations. • Daily spatial coverage of PM 2.5 estimates exceeds 95%, filling in over 70% of the data-missing from satellite AOD. • PM 2.5 distributions indicate severe haze pollution in urban and croplands, as well as in winter. • Potential applications in urban PM 2.5 hotspots location and polluted air mass trajectories capture. |
| ISSN | 0169-8095 |
| Appears in Collections: | 信息系 |
Original Search
