随着中国经济飞速发展，近年来经常出现的大范围严重雾霾污染事件使得空气污染受到政府和公众的广泛关注，成为一个突出的环境问题。空气中空气动力学直径小于等于2.5微米的颗粒物，称为细颗粒物PM_2.5（Particulate matter with Aerodynamic Diameter of 2.5 microns or smaller），是主要的大气环境污染物。流行病学研究表明，PM_2.5浓度与人体不良健康效应存在显著的相关性。中国中东部地区工业和经济发达、人口密集，由于人类活动排放了大量的气溶胶污染物，是细颗粒物污染最为严重的地区之一。卫星遥感能进行大范围、高时空分辨率和多光谱波段的观测，可生成气溶胶光学厚度AOD（Aerosol Optical Depth）产品，通过建立AOD-PM_2.5关系，进而获取近地表PM_2.5浓度。然而，受云遮挡、重污染、卫星轨道等因素影响，卫星反演AOD产品在时间和空间上都存在一定的缺失值，与此同时，PM_2.5质量浓度相关计算模型存在一定的精度问题，这些都严重制约了卫星遥感反演PM_2.5的应用效果。本研究提出基于时空要素极端梯度提升ST_XGBoost（Space-time eXtreme Gradient Boosting）的AOD时空插补方法，并基于该方法获取得到的高精度时空连续AOD数据，建立基于机器学习、综合多要素的中国中东部地区PM_2.5卫星遥感高精度估算方法。论文的主要研究内容和成果如下：

（1）为提高卫星遥感AOD产品时空完整度，提出基于ST_XGBoost的AOD插补方法。该方法综合考虑气象、土地利用类型、时空等要素对AOD的影响，可实现研究区域AOD产品的全覆盖，并与原始AOD产品精度基本保持一致。与MODIS（Moderate resolution Imaging Spectroradiometer）AOD原始产品相比，经插补后的AOD时空完整度从13.2%提升至100%。与气溶胶自动观测网AERONET（Aerosol Robotic Network）AOD相比，决定系数为0.67，均方根误差为0.288，平均偏差为-0.0091。经评估，该方法精度高于现有单个卫星遥感AOD产品的插补精度，也优于大部分多源卫星遥感AOD融合产品的插补精度。分析表明，插补后的AOD北高南低，春季和夏季较高，秋季和冬季较低，2014年～2017年年均值整体呈下降趋势，MODIS AOD的缺失会导致AOD年均值和季节均值偏低。

（2）为减少气溶胶消光系数在垂直分布上呈负指数递减这一假设产生的PM_2.5浓度估算误差，分别提出基于ST_XGBoost和深度信念网络DBN（Deep Belief Network）的两种能见度估算方法。这两种方法综合考虑AOD、气象等因子对能见度的影响，均可一定程度提高能见度估算精度。经对比，研究选取较优的基于ST_XGBoost的能见度估算方法计算能见度，进而直接获取近地面消光系数，提出改进的基于垂直和湿度订正的PM_2.5浓度估算方法。该方法的精度高于传统的基于垂直和湿度订正的PM_2.5浓度估算方法。结果表明，PM_2.5浓度模拟精度从日到年尺度逐渐提升，在年尺度上，决定系数为0.89，均方根误差为5.72μg m^-3； PM_2.5浓度预测精度从日到季节尺度逐渐提升，在季节尺度上，决定系数为0.61，均方根误差为11μg m^-3。

（3）提出一种基于机器学习、综合多要素的中国中东部地区PM_2.5浓度估算方法。该方法利用ST_XGBoost强大的提取数据集本质特征的能力，综合考虑AOD、气象、植被、社会经济和时空等因素对PM_2.5浓度的影响，有效提高PM_2.5浓度估算精度。在利用插补后的AOD进行PM_2.5浓度预测时，基于ST_XGBoost的PM_2.5浓度估算方法精度高于现有方法。该方法模拟精度从日到年尺度逐渐增加，在年尺度上，决定系数为0.92，均方根误差为5.69μg m^-3；预测精度从日到季节尺度逐渐增加，在季节尺度上，决定系数为0.81，均方根误差为10.91μg m^-3。2014年～2016年中国中东部地区基于插补后AOD的PM_2.5年均浓度模拟值为49.50±15.53μg m^-3，2017年PM_2.5年均浓度预测值为48.05±14.44μg m^-3，空间分布均为北高南低，平原高山区丘陵低，内陆高沿海低。对于重点研究区域，基于插补后AOD的PM_2.5年均浓度模拟值和预测值均为华北平原最高，珠三角最低。对于日到季节尺度的模拟和预测精度，基于ST_XGBoost的PM_2.5浓度估算方法均优于垂直和湿度订正方法。

With the rapid development of China's economy, the frequent occurrence of large-scale serious smog and haze pollution events in China during the past years had caused big concern of the government and the public and become a prominent environmental problem. Particulate matter with Aerodynamic Diameter of 2.5 microns or smaller (PM_2.5) is the main atmospheric environmental pollutant. Epidemiological studies had shown that there was a significant correlation between PM_2.5 concentration and human adverse health effects. Central and East China is one of the most developed but heavily PM_2.5 polluted areas, due to the large amount of aerosol pollutants emitted by human activities. With the advantages of wide range, high spatial-temporal resolution and multi spectral bands observation, satellite remote sensing can be used to generate aerosol optical depth (AOD) products. By establishing the AOD-PM_2.5 relationship, PM_2.5 concentration near the surface can be obtained. However, the restrictions of the application and performance of the satellite remote sensing PM_2.5 inversion include the errors of the PM_2.5 concentration models and the missing of the values in time and space of the satellite retrieval AOD products due to the influence of the cloud cover, heavy pollution, satellite orbit and etc. To solve the above issues, in this study, an AOD spatiotemporal gap-filling method based on Space-time eXtreme Gradient Boosting (ST_XGBoost) is proposed to obtain the high-precision spatiotemporal continuous AOD data. Based on these improved AOD data together with the integration of multi-element machine learning, a high-precision remote sensing estimation method for PM_2.5 in Central and East China is established. The main research content and results of the paper are as follows:

In order to improve the space-time integrity of satellite remote sensing AOD products, an AOD gap-filling method based on ST_XGBoost is proposed. This method can realize the full coverage of AOD products in the research area and keep the almost same-level precision with the original AOD products by considering the influence of weather, land use type, time and space. Compared with the Moderate resolution Imaging Spectroradiometer (MODIS) AOD, the space-time integrity of filled gaps of AOD is increased from 13.2% to 100%. Compared with the Aerosol Robotic Network (AERONET) AOD, the determination coefficient，the root mean square error and the bias  are 0.67, 0.288, and -0.0091 respectively. The evaluation demonstrates that the accuracy of this method is higher than that of the existing AOD products of single satellite remote sensing and most AOD fusion products of multi-source satellite remote sensing. The analysis shows that Gap-filled AOD is higher in the north and lower in the south, higher in spring and summer, and lower in autumn and winter. The annual average of Gap-filled AOD during 2014 -2017 shows a downward trend as a whole. The lack of MODIS AOD will lead to the lower annual and seasonal average of AOD.

In order to reduce the estimation error of PM_2.5 concentration caused by the assumption that the extinction coefficient of aerosols decreases negative exponentially in the vertical distribution, two visibility estimation methods based on ST_XGBoost and Deep Belief Network (DBN) are proposed respectively. These two methods can improve the accuracy of visibility estimation accuracy to a certain extent by considering the influence of AOD, meteorology and other factors. After comparison, the study selects the better visibility estimation method based on ST_XGBoost to calculate the visibility, and then directly obtains the near-surface extinction coefficient, and proposes an improved estimation method of PM_2.5 concentration based on vertical and humidity correction. The accuracy of this method is higher than the traditional PM_2.5 concentration estimation method based on vertical and humidity correction. The results show that the accuracy of PM_2.5 concentration simulation gradually improves from the daily to annual scale. On the annual scale, the determination coefficient is 0.89 and the root mean square error is 5.72 μg m^-3; The accuracy of PM_2.5 concentration prediction gradually improves from the daily to seasonal scales. On the seasonal scales, the determination coefficient is 0.61, and the root mean square error is 11 μg m^-3.

The PM_2.5 concentration estimation method for Central and East China is proposed based on comprehensive multi-factors of machine learning. This method makes use of the powerful ability of ST_XGBoost to extract the essential features of data set, comprehensively considers the influence of AOD, meteorology, vegetation, socio-economic and space-time on PM_2.5 concentration, and effectively improves the precision of PM_2.5 concentration estimation. When using Gap-filled AOD to predict the PM_2.5 concentration, the accuracy of the method based on ST_XGBoost is higher than that of the existing method. The simulation accuracy of this method gradually increases from the daily to annual scales. On the annual scale, the determination coefficient is 0.92 and the root mean square error is 5.69μg m^-3; the prediction accuracy gradually increases from the daily to seasonal scales. On the seasonal scale, the determination coefficients is 0.81, and the root mean square error is 10.91μg m^-3. The simulated annual average value of PM_2.5 concentration based on Gap-filled AOD from 2014 to 2016 in Central and East China is 49.50 ± 15.53μg m^-3, and the predicted average annual PM_2.5 concentration in 2017 is 48.05 ± 14.44μg m^-3.The spatial distribution is higher in the north, mountains, hills and inland than that in the south, in the plains, in the hills and in the coast respectively. For key research areas, the simulated and predicted annual average PM_2.5 concentration based on Gap-filled AOD is the highest in the North China Plain and the lowest in the Pearl River Delta. For the simulation and prediction accuracy on a daily to seasonal scale, the PM_2.5 concentration estimation method based on ST_XGBoost is superior to the vertical and humidity correction methods.