【作者】 刘宁；

【导师】 邹滨；

【作者基本信息】 中南大学 ， 摄影测量与遥感， 2022， 博士

【摘要】 我国空气质量近年虽在逐步改善,但局部区域的重污染天气仍时有发生,其中以PM_2.5为首要污染物的污染事件尤为频繁。厘清重点区域PM_2.5生消扩散的时空演化规律、并提前准确预测未来时刻PM_2.5浓度时空分布,对我国大气污染防控治理依然十分重要,因此实现PM_2.5浓度实时无缝精细模拟与预报是当前空气污染防控治理的现实需求。对此,本论文拟以京津冀城市群及周边区域为研究区,综合考虑多模态数据优势互补特性,耦合高精度的地基监测数据、高空间/时间分辨率的极轨/静止卫星反演数据、日夜间实时无缝的数值模式模拟与预报数据,开展多模态数据融合的PM_2.5浓度实时无缝精细模拟与预报研究。在充分验证多模态AOD/PM_2.5产品精度后,构建了多模态数据融合下日夜间全时段AOD实时无缝精细优化模型,发展一套支撑小时尺度PM_2.5浓度高效动态估算的LMESTKF模型,最终耦合数值模式实现高精度PM_2.5浓度时空分布精细预报。研究总结具体如下:（1）基于高精度地基监测数据评估了遥感反演及数值模式等多模态数据精度。选取1 km MAIAC与750 m VIIRS＿IP高分辨率极轨卫星AOD产品、5 km Himawari-8与6 km GOCI的小时级静止卫星AOD产品、0.25°/3 h GEOS-FP与0.625°×0.5°/1 h MERRA2的数值模式AOD/PM_2.5再分析/预报产品作为多模态数据融合源,以高精度的地基PM_2.5浓度与地基AERONET AOD数据为真值,在统一验证标准下开展卫星遥感反演与数值模式产品精度验证研究,为后续多模态数据融合提供数据筛选与精度定权依据。主要结果表明,MAIAC与GOCI AOD产品精度最高（R²/RMSE在0.87/0.15左右）、Himawari-8 AOD产品精度次之（R²/RMSE为0.67/0.20）、VIIRS＿IP AOD精度较差（R²/RMSE仅为0.43/0.35）。GEOS-FP/MERRA2的AOD再分析产品精度（R²/RMSE为0.6/0.4）略低于遥感反演结果,而GEOS-FP/MERRA2的PM_2.5再分析产品验证精度较差（R²/RMSE仅为0.15/50 g/m³）。相比再分析产品而言,GEOS-FP/MERRA2的AOD/PM_2.5预报产品在预报初始时刻精度与再分析产品类似,但随预报时长的增加精度呈递减趋势。（2）构建了以数值模式填补遥感反演缺失值的日夜间全时段实时无缝AOD优化模型。充分利用数值模式AOD日夜间无缝模拟优势,采用精度验证结果对多模态数据定权的基础上,基于随机森林（Random Forest,RF）与自编码器的深度残差网络（Auto Encoder-based Residual Network,Auto Res Net）融合多模态数据,成功解决了遥感AOD产品中大量缺失及夜间无法反演的问题,生成了日夜间全时段的1 km/1 h分辨率无缝AOD数据。经地基AERONET AOD验证发现,在没有遥感数据覆盖的区域,AOD融合结果精度略低于有遥感数据区域,但验证R²、RMSE、Within＿EE与Bias指标分别仍可达0.83、0.21、62.50%与0.03左右;相比AOD融合结果与地基PM_2.5浓度的相关性在日间与夜间的差异,二者相差不大,说明夜间AOD融合结果对PM_2.5浓度的解释力与日间结果处在同一水平。相比于以往数值模式填补遥感AOD的研究,本研究不仅考虑了多颗卫星反演产品,同时能实现AOD日夜间全时段小时级的精细融合。（3）提出了高效的小时级PM_2.5浓度时空模拟模型（LME and Spatiotemporal Kalman Filter,LMESTKF）。该模型耦合传统LME模型与高效时空Kalman模型解决了传统小时级AOD-PM_2.5浓度估算的时空统计模型耗时问题,同时借助站点留一交叉验证的思想,提出了LMESTKF空间降维算法,进一步提升LMESTKF的建模效率、降低过拟合现象。基于LMESTKF模型,建立了地基PM_2.5浓度与1km/1 h分辨率无缝AOD的统计关系,实现了日夜间全时段的1 km/1h分辨率无缝PM_2.5浓度的估算。结果表明,基于样本与站点的十折交叉验证R²、RMSE、Bias与MAPE指标分别为（0.91、14.37 g/m³、-0.41 g/m³、28.91%）与（0.87、16.98 g/m³、-0.32 g/m³、34.62%）,其精度高于现有京津冀城市群PM_2.5浓度估算模型;此外构建月尺度下日夜间小时级PM_2.5浓度估算的LMESTKF模型所需的时间消耗与内存消耗也分别仅在13.85分钟、1.67 GB左右,说明本研究提出来的LMESTKF模型不仅精度高,效率也快。相比于传统PM_2.5浓度时空模拟研究多在日尺度、月尺度及年尺度,本研究的所提出来的LMESTKF模型可实现小时尺度PM_2.5浓度时空高效模拟。（4）提出了耦合数值模式预报的未来5天高精度PM_2.5浓度实时无缝精细统计预报框架。该框架在耦合数值模式预报的基础上,分别基于RF与自编码器的深度卷积残差网络（Auto Encoder-based CNN Residual Network,Auto CNNRes Net）实现高精度地基PM_2.5浓度时序预报及AOD实时无缝精细预报,进一步针对污染成因机制稳定性采用站点同化的历史模型迁移预报策略或未来情景的直接建模预报策略,实现未来5天高精度1 km/3 h分辨率PM_2.5浓度实时无缝精细预报。相应精度验证结果显示,在临近时刻的预报结果验证R²、R、RMSE、Bias分别为0.80、0.90、9.71 g/m³、0.76 g/m³。与传统数值模式预报结果比较,本论文的预报结果验证R²、R与RMSE的精度提升比例分别为79.59%、48.12%及64.50%。但PM_2.5浓度实时无缝预报结果精度随预报时长的增长呈现下降趋势,在第120小时的预报结果验证R²、R、RMSE、Bias下降至0.30、0.57、18.11g/m³、-0.05 g/m³。相应信息熵评估结果显示,本论文的预报结果相比于数值模式预报结果空间信息更加丰富,信息熵提升比例最高可达140.60%、平均提升36.92%。相比于传统的PM_2.5浓度数值预报方法,本研究所提出的方法预报精度更高、空间信息更加丰富。综上所述,针对现有PM_2.5浓度模拟时空缺失严重、预报结果粗糙不准等问题,本论文研究成果实现了日夜间无缝公里小时级AOD/PM_2.5估算,创建的LMESTKF估算模型可提高PM_2.5浓度实时无缝精细模拟的精度与效率,提出的高精度PM_2.5浓度精细预报框架可进一步提升我国大气污染防控治理的水平。图110幅,表22个,参考文献233篇更多还原

【Abstract】 Although China’s air quality has been gradually improved in recent years,heavy air pollution events still occur in some local regions,and pollution incidents with PM_2.5 as the primary pollutant are particularly frequent.It is still very important to clarify the spatiotemporal evolution rules of the generation,elimination and diffusion of PM_2.5 concentration in key regions,and to accurately predict the spatiotemporal distribution of PM_2.5 concentration in the future.Therefore,it is the real needs of current air pollution prevention and control to realize the seamless fine simulation and forecast of real-time PM_2.5 Concentration.In this regard,this paper intends to take the Beijing-Tianjin-Hebei urban agglomeration with surrounding areas as the research area,comprehensively consider the complementary characteristics of multi-modal data,and couple ground-based monitoring data with high accuracy,polar-orbiting and geostationary satellite inversion data with high spatial or temporal resolution,numerical model data with real-time seamless simulation and forecast in all periods of day and night,and carry out the research on seamless fine simulation and forecast of real-time PM_2.5 concentration using mutli-modal data fusion.Specifically,after fully evaluating the accuracy of multi-modal AOD/PM_2.5 products,this paper fuses the multi-modal data to realize seamless fine model of real-time AOD in all periods of day and night,and develops a LMESTKF model to support efficient estimation of PM_2.5concentration at hourly scale,and couples numerical model forecasting results to achieve the spatiotemporal fine forecast of PM_2.5 concentration distribution with high accuracy.The research summary is as follows:（1）Validate the accuracy of multi-modal data such as satellite-based inversion and numerical models using high-accurate ground-based monitoring data.Select 1 km/1 day MAIAC and 750 m/1 day VIIRS＿IP AOD products,5 km/1 h Himawari-8 and 6 km/1 h GOCI AOD products,0.25°/3 h GEOS-FP and 0.625°×0.5°/1 h MERRA2 AOD/PM_2.5reanalysis/forecast product of numerical model as multi-modal data fusion source,and take the ground-based PM_2.5 concentration and AERONET AOD data as the true value,evaluate the accuracy of multi-modal data using the unified validation criterion.The main results show that the MAIAC and GOCI AOD products achieve the highest accuracy（R²/RMSE is around 0.87/0.15）,the Himawari-8 AOD product obtains the second highest accuracy（R²/RMSE is 0.67/0.20）,and the VIIRS＿IP AOD has poor accuracy（R²/RMSE is only 0.43/0.35）.The AOD reanalysis accuracy of GEOS-FP/MERRA2（R²/RMSE is 0.6/0.4）products is slightly lower than the satellite-based inversion results,and the PM_2.5 reanalysis accuracy of GEOS-FP/MERRA2 products is poor（R²/RMSE is only 0.15/50 g/m³）.Compared with the reanalysis products,the AOD/PM_2.5 forecast results of GEOS-FP/MERRA2 products have similar accuracy to the reanalysis products at the initial forecast time,but the accuracy decreases with the increase of the forecast time.（2）Construct a real-time seamless AOD optimization model for all periods of day and night by filling the missing data in satellite-based products with numerical model results.Taking the seamless simulation advantage of in all periods of day and night in the numerical model AOD product,and using the accuracy validation results to weight the multi-modal data,an autoencoder-based deep residual network and random forest model is employed to fuse multi-modal data which successfully solves the problem of large missing gaps in satellite-based AOD products and generates seamless AOD data with 1 km/1 h resolution for the whole period of day and night.The validation results based on ground-based AERONET AOD find that the accuracy of AOD fusion results in areas without satellite data is slightly lower than that those areas with satellite data,but the validated R²,RMSE,Within＿EE and Bias in areas without satellite data can still reach about 0.83,0.21,62.50%and 0.03 respectively.The correlation between AOD fusion results and ground-based PM_2.5concentration shows no obvious discrepancy between daytime and nighttime which indicates that the explanatory power on PM_2.5concentration of AOD fusion results at nighttime keeps the same level with the fusion results at daytime.Compared with previous studies on filling remote sensing AOD with numerical models,this study not only considers the retrieval products of multiple satellites,but also realizes the fine fusion of hourly AOD at all period of the day and night.（3）Propose an efficient spatiotemporal dynamic simulation model（LME and Spatiotemporal Kalman Filter,LMESTKF）of hourly PM_2.5concentration.The model couples the traditional LME model and the efficient spatiotemporal Kalman model to solve the time-consuming problem of the traditional spatiotemporal statistical model for AOD-PM_2.5concentration estimation.Using the idea of cross-validation,a spatial dimensionality reduction algorithm of LMESTKF is also proposed.Based on the LMESTKF model,the statistical relationship between ground-based PM_2.5 concentration and seamless AOD data at 1 km/1 h resolution is established,the seamless PM_2.5 results at 1 km/1 h resolution are also generated.The results show that the R²,RMSE,Bias and MAPE of the LMESTKF model based on sample-based and site-based cross-validation are（0.91,14.37 g/m³,-0.41 g/m³,28.91%）and（0.87,16.98 g/m³,-0.32 g/m³,34.62%）,the accuracy of the proposed LMSETKF model is higher than the existing PM_2.5 concentration estimation model at Beijing-Tianjin-Hebei region.In addition,the time and memory consumption of the LMESTKF model for hourly PM_2.5 concentration mapping at the monthly scale is only about 13.85 minutes and 1.67 GB respectively,which shows that the LMESTKF model proposed in this study not only obtains higher accuracy but also achieves faster efficiency.Compared with traditional spatiotemporal simulation studies of PM_2.5 concentration,which are mostly conducted on the daily,monthly and annual scales,the LMESTKF model proposed in this study can achieve efficient spatiotemporal simulation of hourly PM_2.5 concentration.（4）Propose a high accurate statistical forecasting framework of real-time seamless and fine PM_2.5 concentration in the next 5 days by coupling with numerical model forecasting results.On the basis of coupled numerical model forecast,the framework firstly realizes ground-based PM_2.5 concentration time series forecast with high accuracy and AOD real-time forecast with fine resolution using RF model and autoencoder-based CNN residual network respectively.Then,considering the stability of the pollution cause mechanism,the framework also adopts the historical model migration forecasting strategy with site assimilation or the direct modeling forecasting strategy of future scenarios to achieve real-time seamless forecast of high-accurate fine PM_2.5 concentration with 1 km/1 h resolution in the next 5 days.The corresponding accuracy validation results show that the R²,R,RMSE,and Bias of the forecasting results at nowcasting time are 0.80,0.90,9.71 g/m³,and 0.76 g/m³,respectively.Compared with the traditional numerical model forecast,it is found that the forecast results based on the proposed forecast framework improves the R²,R and RMSE at 79.59%,48.12%and 64.50%ratio,respectively.However,the accuracy of the proposed forecasting framework showed a downward trend with the forecast time.On the fifth day,the R²,R,RMSE,and Bias of the forecast results drops to 0.30,0.57,18.11 g/m³,-0.05 g/m³.The information entropy evaluation results also point out that the forecasting results based on the proposed framework are more abundant in spatial information than the numerical model forecasting results,and the information entropy improvement ratio can reach up to 140.60%,with an average increase of36.92%.Compared with the traditional numerical prediction method of PM_2.5 concentration,the method proposed in this study has higher prediction accuracy and richer spatial information.In summary,regarding to the problems of the existing model which simulation results contain largely spatiotemporal missing gaps and forecast results are rough and inaccurate,the research results of this paper have realized the seamless hourly AOD/PM_2.5 estimation at kilometric scale in all periods of day and night,and the created LMESTKF model can improve the accuracy and efficiency of seamless fine simulation of real-time PM_2.5concentration,and the corresponding high-accurate PM_2.5 concentration fine forecast framework can further improve the level of air pollution prevention and control in China.更多还原