异构时空深度学习法评估高时空分辨率空气质量

Against the missing values of the covariates in assessment of air pollutants, and insufficient consideration of spatial autocorrelation in the existing models, this study proposes an approach of heterogeneous spatiotemporal deep learning to high-accuracy estimation of the air pollutants (PM2.5 and NO2) concentrations at high spatiotemporal resolution, and to scenario simulation and optimal decision under the climate changes. In the aspect of data completeness, this project will employ auto-encoder to extract the feature representation and utilize the symmetry network to interpolate the missing covariates and reduce the over-fitting. This project will use the convolutional layers, kernel density and the nearest-neighbors layer to model spatial autocorrelation, and proposes the stratified heterogeneous deep network models with combination of the convolutional and other heterogeneous layers. We utilize the flexible network structure and larger parameter space to comprehensively capture spatial autocorrelation and the non-linear associations, thus reducing the bias and improving spatiotemporal point prediction accuracy. For the missing estimates due to missing covariates, we propose the constrained optimal models based on the high-accuracy point estimates and temporal basis functions to capture seasonal patterns of air pollutants and improve spatiotemporal coverage of the predictions. Subsequently, this project will simulate spatiotemporal distributions of air pollutants in the future under the context of climate changes and evaluate the different measures against air pollution under different scenarios to conduct the optimal decision-making support. This study will compensate the shortcomings of insufficient spatiotemporal coverage, reduce the bias in estimation, and provide valuable decision-making support information for the national measures against air pollution under the effect of climate changes.

针对空气污染物浓度评估中协变量缺失，已有模型空间相关性考虑不足，本研究提出异构时空深度学习法，预估高精度高时空分辨率空气污染物（以PM2.5及NO2为例）浓度，进行气候变化情景模拟及优化决策。在数据完整性方面，课题拟采用自编码方法提取特征表征，利用网络对称结构插补协变量，减少过拟合；课题采用卷积层、核密度及最近邻层建模空间相关性，提出将卷积层同普通层融合的异构深度网络，利用更灵活网络结构与更大参数空间全面捕捉空间自相关性及非线性关系，减少偏差，提高时空点预测精度；针对协变量缺失导致估计缺值，提出基于高精度点估计及时间基函数的限制性优化模型，捕捉其季节性规律，提高预测时空覆盖度。在此基础上，课题预估气候变化下空气污染未来时空分布，模拟不同应对措施，进行优化辅助决策。本研究将弥补当前空气污染监测及评估的时空覆盖度的不足，减少评估偏差，为国家在气候变化下应对空气污染提供有价值辅助决策信息。

针对空气污染物浓度评估中协变量缺失，已有模型空间相关性考虑不足等，本项目研建了异构时空深度学习系统性建模方法，全面提高空气质量时空预测的分辨率及泛化性。研究提出了高精度气象参数反演模型，特别是设计了空间权重回归同深度学习结合的风速预测取得了高可靠的输出，为后续缺值插补及空气质量评估提供了关键气象数据；卫星数据的高精度全覆盖插补深度学习方法，在MAIAC AOD及OMI-NO2测试插补精度（R2）达到0.9以上；探索了空间相关性的深度建模方法，完善了限制性优化时空趋势模拟并完成不确定性评估，实现在稀疏样本地区时空全覆盖的趋势模拟；建立了将图网络同残差结合的图混合深度建模方法，耦合稀疏样本同时空大数据提高模型预测泛化性（独立测试R2：0.84（NO2）及0.89（PM2.5））。数据方面，完成了从2015年到2018年的中国大陆地区的MAIC AOD及OMI-NO2的卫星遥感气溶胶缺失数据的全覆盖插补，提高这些卫星数据产品的可用性；完成关键气候变量、主要空气污染物及空气质量指数的高精度高时空分辨率参数反演；评估了4年空气质量变化对健康效应的影响，重点评估了对非意外死亡的影响。研究提出全残差深度回归建模方法，模型易于训练，可获得高精度预测结果，该方法具有广泛应用潜力，研究成果被机器学习、遥感、资源环境等领域国际专业顶刊等引用。估算的空间空气污染浓度在公共健康影响评估中取得很好效果，在美国胸科学会年报中指出提高了估计结果同健康效应关系，更正了气喘同NOx的统计关联，在确立NOx同胸腺癌间关联上提供关键污染数据支持。