空间贝叶斯方法及在空气质量及其健康效果评估中的运用

The existing methods of evaluating air quality and its health effects are mainly based on spatial interpolation, generalized additive model (GAM), logistic regression or land-use regression (LUR). The shortcomings of these traditional methods include: (1) consideration of few predictive factors in spatial interpolations such as kriging or no spatial correlation in GAM and LUR; (2) linear regression mostly used to establish the association between predictive factors and exposure or health effects; (3) difficulty of combination and explanation with a priori knowledge; (4) use of a limited number of samples to train the models, probably resulting in biased results. Under this context, this study proposes spatial Bayesian approaches that are based on the Bayesian theorem, incorporate the information of spatial clustering or spatial correlation and employ additional linear/non-linear spline functions to establish the network links or regression associations. Our Bayesian approaches are able to employ a priori knowledge to improve justification of the results, take full consideration of spatial confounding effect to reduce the bias in estimation, and adopt additional terms to establish the non-linear association between predictive factors and the target variable (exposure to air pollutants and health effect) to reflect the "dose-response" relationship in practice. The parallel version of the algorithms will be also realized to deal with massive datasets to improve the computation efficiency and estimation accuracy; the scenario of exposure to air pollution and the corresponding effects for the climate changes in the future will also be simulated and the output will provide informative support for the relevant policies of air quality.

国内外常用的空气质量及健康影响评估包括克里格插值、概化累加模型（GAM）、逻辑斯特回归及采用土地利用变量的回归（LUR）。传统方法主要不足：空间估值方法考虑辅助变量较少，GAM及LUR模型没有考虑空间相关性，LUR及逻辑斯特回归基于线性关系，与机理结合较薄弱，样本过少，可能会导致有偏结果。针对已有方法不足，本研究提出空间贝叶斯方法，该方法基于贝叶斯原理，加入空间聚类或空间相关性因素，引入累加光滑函数建立模型。该方法将结合机理或已有文献获取先验知识，充分考虑空间混淆因素影响，减少偏差，提高估计精度；采用累加模型提取影响因素同目标变量（暴露度/健康效果）之间线性/非线性关系，在健康效果评估中更符合实际地量化"剂量-反应"关系。算法将对相关海量数据进行处理，提高提高评估效率；对未来气候变化情景进行效果模拟，为空气质量政策提供信息支持。

本项目针对已有空气污染监测点密度低导致评估精度不高及空气污染健康效果评估中空间相关性考虑不足，提出了空间贝叶斯方法，以提高空气污染暴露度估计精度及区域化健康效应的评估。本项目执行顺利，已完成项目规定的各项任务，取得有影响力的成果。围绕着空间贝叶斯的环境暴露度评估及健康风险分析的总目标，课题研制了：（1）离散空间贝叶斯模型，该模型将空间/时空聚集信息引入贝叶斯网络，通过空间推理建立空间贝叶斯网络，应用到健康风险评估中减少了评估中未考虑空间自相关导致的偏差；（2）三阶段限制性混合非线性时空预测模型，课题首次提出基于泰森多边形的时空模型捕捉目标变量的时空变异及时空聚集机器学习，提高估值稳定性及精度，限制性优化算法可在有限实测或/及高精度点估计情况下模拟全时间段内时间序列，弥补传统模型由于协变量缺失导致预测亦缺失的不足，该三阶段模型在NO2及NOx的均取得较高精度；（3）PM2.5时空浓度评估模型，基于天的三阶段模型的架构，避免采用单一模型混淆不同时段的空间效应及变量间关系的缺陷而大幅提高建模精度；（4）连续空间贝叶斯健康风险评估区域模型，融入空间相关性及非线性，有效捕捉了社区因素对空气污染的健康效果的影响；（5）基于自解码的残差深度网络，提高气象预测中的学习效率。基于扎实的方法理论基础，课题研制了支持并行化的空间贝叶斯算法及空气污染时空评估模型软件。后者简化了时空建模的各项空间运算操作，通过并行支持海量基模型的学习，成倍提高效率。算法通过R及Python开源包共享，对于推广应用本项目提出的方法及相关学科的发展具有重要实际意义及应用价值，为课题组继续制作覆盖全国的高精度高分辨率的空气污染栅格提供极有价值的参考。本课题共发表文章12篇，其中SCI文章9篇（2篇1区论文），中文核心文章2篇；发布软件包4个，其中3个开源；申请公示专利2项，获得专利授权1项；指导完成3篇硕士生论文。