面向场景的城市PM2.5浓度空间分布精细模拟

针对传统PM_2.5浓度空间分布模拟方法忽略了城市内部如道路、工厂、居民区、景区等不同微环境整体对PM_2.5浓度影响机制的缺陷，本研究提出一种微环境PM_2.5浓度场景分异的理论假设，并以湖南长沙主城区为例，结合基于污染先验知识划分的城市微环境场景空间分布与自主设计加密观测场获取的203个监测点小时PM_2.5浓度加密数据，分析城市微环境PM_2.5浓度场景时空分异特征.在此基础上，耦合地理加权回归（GWR）与人工神经网络（ANN）方法，构建微环境场景增强下的PM_2.5浓度空间分布精细模拟GWR-ANN模型，开展城市内部高空间分辨率PM_2.5污染制图.结果表明：不同微环境场景间PM_2.5浓度存在显著时空差异，地表覆盖类型相同但分别位于2个不同场景的监测点间PM_2.5浓度差会随时间发生变化；耦合微环境场景变量的GWR-ANN模型能够有效精细模拟PM_2.5浓度的空间分布，模型拟合效果与交叉检验精度指标整体优于无场景变量参与的GWR-ANN模型（除部分时相较为接近外，检验R²：0.76~0.84vs.0.57~0.81）；场景增强下的PM_2.5浓度空间分布100m级分辨率模拟估算结果可以较好揭示研究区PM_2.5浓度高低值局地变化特征.

Urban scene-oriented simulation of the spatial distribution of PM2.5 concentration in an intra-urban area at fine scale

The traditional spatial simulation technologies of PM_2.5 concentration usually ignored the mechanism behind the PM_2.5-urban scenes (e.g. roads, factory, residential area, scenic area) correlation. This study proposed an urban scene assumption of PM_2.5 concentration, namely the PM_2.5 concentration is rather homogeneous within an urban scene while heterogeneous between different urban scenes. Taking the intra-urban area of Changsha, Hunan as an example, the spatial distribution of urban scenes was manually interpreted using a priori knowledge and a high-density monitoring sampling campaign was conducted for two periods in December 24~25, 2015. Based on the hourly PM_2.5 concentration observations from 203 sampling sites and the urban scene map, the urban scene difference of PM_2.5 concentration was explored and an urban scene enhanced two-stage modelling strategy of geographically weighted regression and artificial neural networks (GWR-ANN) was developed. The spatial patterns of PM_2.5 concentrations were simulated based on GWR-ANN at the 100×100m resolution. Results show that the spatiotemporal variations of PM_2.5 concentration between urban scenes do exist and the difference of PM_2.5 concentration for sampling sites with the same land use/cover in two different types of urban scenes varied with time. The urban scene enhanced GWR-ANN could be effective in spatial simulation of PM_2.5 concentrations at fine scale. The GWR-ANN model with urban scene variable performed better than the GWR-ANN model without urban scene variable. Except for five sampling hours with rather close statistics, the cross-validation R² between estimated PM_2.5 concentration and observed PM_2.5 concentration for GWR-ANN with urban scene were higher than GWR-ANN without urban scene (0.76~0.84vs. 0.57~0.81). The spatial patterns of PM_2.5 concentrations based on urban scene enhanced GWR-ANN could be effective in disclosing detail hot-spots and cold-spots of PM_2.5 pollution.