面向城市微环境场景的PM2.5浓度空间分布精细模拟

Monitoring urban PM2.5 concentrations at a fine spatial scale in which individuals live and work has been becoming a new demand for the decrease of the national public health risk. However, the widely used advanced statistical modeling at home and aboard for fine PM2.5 concentration monitoring is still problematic because of the sparse monitoring stations and the defective modeling theory hypothesis. In view of this, based on the transfer learning idea, this project will for the first propose and validate a new spatial-temporal heterogeneity theory hypothesis through the densified observation in the source regions (i.e. typical microenvironment scenarios in urban area), and consequently reveal the geographical law of spatial-temporally differentiated PM2.5 concentrations at the urban microenvironment scenarios. Meanwhile, scenario configuration-oriented deep learning intelligence model for PM2.5 concentration modeling will be developed. And the inner mechanism for PM2.5 concentration intelligence model transferring between source regions and target regions will be claimed under the conditions of similar scenario configurations. In summary, as an innovative attempt, this project will realize the PM2.5 concentration mapping at fine-resolution in large target regions with sparse monitoring stations through the densified observation samples in source regions. The newly proposed spatial-temporal heterogeneity theory hypothesis, as well as the consequently developed deep learning intelligence modeling and transferring methods in this project will provide a new theoretical and method supports for fine scale mapping of urban air quality, and enhance the levels of national techniques for microenvironment scenario-oriented air pollution monitoring and prevention.

精细监测城市人居活动场景PM2.5浓度是新形势下国家全面降低空气污染健康损害的新需求。但国内外现广泛采用的高级统计建模监测手段在城市PM2.5浓度空间分布精细模拟中仍普遍受到监测站点数量与建模理论假设缺陷限制。依托迁移学习思想，项目拟通过在城市内部典型场景（源领域）开展加密监测，提出与实证全新的PM2.5浓度时空分异理论假设，揭示城市微环境场景PM2.5浓度时空分异的地理规律；研究建立面向场景配置的PM2.5浓度深度学习智能模型，阐明场景配置相似下的源领域PM2.5浓度智能模型向稀疏监测目标领域迁移的内在机制。作为“以在源领域小范围补充加密建模样本、实现目标领域大范围PM2.5浓度空间分布精细模拟”的创新尝试，项目提出的微环境场景PM2.5浓度时空分异理论假设、研发的空间分布智能建模与迁移方法，可为城市空气质量时空精细模拟制图提供新的理论与方法支撑，提升我国大气污染空间监管与防控技术水平。

掌握大气污染时空精细分布是我国现今开展大气污染精准防控的现实需要，也是全面降低公众大气污染健康损害的重要需求。然而，地面定点监测、卫星遥感监测、移动走航监测等手段目前仍然无法有效实现政府/公众对特定场景空气质量的精准认知。为此，本项目围绕“城市微环境场景PM2.5浓度时空分异的地理规律”和“源领域PM2.5浓度智能模型向目标领域迁移的内在机制”两大科学问题，探索了通过源领域（小范围）的加密观测实现城市稀疏监测目标领域（大范围）PM2.5浓度精细模拟的方法。具体而言，开展了城市PM2.5浓度时空变化的微环境场景依赖分析、PM2.5浓度微环境场景的地理理解与城市全域划分、加密监测典型微环境场景PM2.5浓度智能建模、微环境场景配置相似下的城市全域PM2.5浓度模拟验证及场景化应用等研究。研究发现，不同微环境场景间的PM2.5浓度存在显著时空差异，且与土地覆盖、开发强度、景观格局等因素高度关联。轻度污染下，城市商业区场景的PM2.5浓度均值一般高于其他场景类型；重度污染下，城市PM2.5浓度均值最高和最低场景类型通常为建筑区与景区。面向微环境场景的GWR-ANN模型可在百米级分辨率较好模拟城市PM2.5浓度的空间分布（验证R2：0.76-0.84），并在社区尺度空气污染迁移预测中较为稳健，拟合R2最高可达0.966。本项目提出的面向微环境场景的大气污染时空分异理论假设突破了城市PM2.5浓度时空精细智能模拟的理论瓶颈，为有限监测数据条件下实现城市大范围、高分辨率、高精度的空气质量制图提供了一套全新的技术手段，可助力城市大气污染监管巡查业务中的热点网格目标锁定能力由3km范围精细到100m。项目成果作为关键技术入选了生态环境部全国16个城市大气污染智慧监测试点长沙项目，同时为湖南区域减污降碳时空精准协同治理提供了强有力的理论与方法支撑，获得省长、副省长等领导批示。

内容获取失败，请点击重试