基于三阶段深度学习时空填补模型的京津冀气溶胶光学厚度全覆盖研究
- Comprehensive Study of Regional Aerosol Optical Thickness in the Beijing-Tianjin-Hebei Region Using a Three-Stage Deep Learning Spatio-Temporal Imputation Model
- 基金项目：国家重大科研仪器研制项目(No.42327806)；浙江省“领雁”研发攻关计划(No.2022C03073)；重庆市自然科学基金面上项目(No.CSTB2023NSCQ-MSX0305)；重庆市教委科学技术研究项目(No.KJZD-M202201402)
- 摘要：气溶胶对人体健康危害显著，监测其连续分布范围意义重大。随着卫星遥感技术的发展以及算法的优化，利用卫星遥感数据可以获得连续分布的气溶胶数据。基于多角度大气矫正算法获得的气溶胶光学厚度数据(MAIAC AOD)具有高时空分辨率，对于评估区域气溶胶水平以及人体暴露风险评估具有重要意义。然而，地表反射和云覆盖等因素导致2021至2023年京津冀地区MAIAC AOD的平均覆盖率仅为42.57%。在本研究，为了解决以往单一模型无法准确的回归不同缺失程度MAIAC AOD的问题，提出了一种基于三阶段深度学习的时空填补方法，先后使用卷积生成、多元回归、时序预测三种回归模式对MAIAC AOD不同覆盖面积缺失数据进行填补，取得了较好填补的结果。经过三阶段填补的MAIAC AOD数据与全球气溶胶观测网络(AERONET)气溶胶光学厚度数据有很好的相关性(R2 = 0.87; RMSE = 0.058)，以上结果表明本填补方法可以提供可靠的全覆盖AOD估计，并为下游PM2.5等空气质量模型提供高质量的数据集。
- Abstract：Aerosols have a significant negative impact on human health, and monitoring their continuous spatiotemporal distribution is of great importance. With the development of satellite remote sensing technology and the optimization of algorithms, it can be obtained that continuous spatiotemporal distribution of aerosol from satellite remote sensing. The MAIAC AOD, based on the multi-angle atmospheric correction algorithm, has high spatiotemporal resolution and is great helpful to estimate aerosol level and human exposure risk to aerosol in a region. However, some factors such as surface reflection and cloud cover have caused in the great disappearance of MAIAC AOD in the Beijing-Tianjin-Hebei region from 2021 to 2023 with an average coverage of only 42.57% .. In this study a three-stage deep learning spatiotemporal imputation method was developed to overcome the drawback that single model could not well regress MAIAC AOD with missing data,. The missing data were successfully recovered using three algorithms including convolutional generation, multiple regression, and time series prediction, respectively. The MAIAC AOD data after the three stages of recovery showed a good correlation with the Aerosol Robotic Network (AERONET) aerosol optical thickness data (R2 = 0.87; RMSE = 0.058). This work indicates that the imputation method can provide reliable full coverage AOD estimates and provide a high-quality dataset for air quality models.
