地球信息科学学报 ›› 2020, Vol. 22 ›› Issue (1): 122-135.doi: 10.12082/dqxxkx.2020.190533

• 专辑:地理智能 • 上一篇    下一篇

基于地理神经网络加权回归的中国PM2.5浓度空间分布估算方法

杜震洪1,2,*(), 吴森森1,2, 王中一1, 汪愿愿1,2, 张丰1,2, 刘仁义1,2   

  1. 1. 浙江大学地球科学学院,地理与空间信息研究所,杭州 310027
    2. 浙江省资源与环境信息系统重点实验室,杭州 310028
  • 收稿日期:2019-09-22 修回日期:2019-10-27 出版日期:2020-01-25 发布日期:2020-04-08
  • 通讯作者: 杜震洪 E-mail:duzhenhong@zju.edu.cn
  • 基金资助:
    国家自然科学基金项目(41871287);国家自然科学基金项目(41701436);国家重点研发计划项目(2018YFB0505000);中央高校基本科研业务费专项资金(2019QNA3013)

Estimating Ground-Level PM2.5 Concentrations Across China Using Geographically Neural Network Weighted Regression

DU Zhenhong1,2,*(), WU Sensen1,2, WANG Zhongyi1, WANG Yuanyuan1,2, ZHANG Feng1,2, LIU Renyi1,2   

  1. 1. School of Earth Sciences, Zhejiang University, Hangzhou 310027, China
    2. Zhejiang Provincial Key Laboratory of Geographic Information Science, Hangzhou 310028, China
  • Received:2019-09-22 Revised:2019-10-27 Online:2020-01-25 Published:2020-04-08
  • Contact: DU Zhenhong E-mail:duzhenhong@zju.edu.cn
  • Supported by:
    National Natural Science Foundation of China(41871287);National Natural Science Foundation of China(41701436);National Key Research and Development Program of China(2018YFB0505000);Fundamental Research Funds for the Central Universities(2019QNA3013)

摘要:

中国空气污染问题日益严重,为获得连续的PM2.5浓度空间分布,现有研究建立了多种基于统计回归的PM2.5估算模型。然而,由于PM2.5回归关系显著的空间非平稳性和复杂的非线性特征,如何实现高精度、高合理性的PM2.5浓度空间大面估计仍然面临挑战,尤其在地形变化复杂、覆盖范围广阔的中国地区更为突出。本文引入了一种将普通线性回归(OLR)和神经网络结合的地理神经网络加权回归(GNNWR)模型,通过集成遥感数据、气象数据和地理信息数据建立了基于GNNWR的PM2.5浓度空间估算方法。文章以中国2017年PM2.5年平均浓度估算为例,开展了该模型与OLR、地理加权回归(GWR)的比较实验。实验结果表明,基于GNNWR的PM2.浓度估算性能指标均明显优于OLR和GWR,且预测精度显著高于GWR。此外,GNNWR获得的PM2.5浓度空间分布也更为合理,较为细致地刻画了中国地区PM2.5浓度的局部空间变化和细节层次。

关键词: 空气污染, PM2.5浓度估算, 空间非平稳性, 地理神经网络加权回归, 地理加权回归, 普通线性回归, 遥感观测, 中国

Abstract:

China is becoming one of the most air-polluted countries and is experiencing severe PM2.5 pollution. To acquire spatially continuous PM2.5 estimates, numerous statistical methods have been developed through the integration of ground-level measurements and satellite-based observations. The estimation of PM2.5 concentrations in China is characterized by significant spatial nonstationarity and complex nonlinearity due to the complicated terrain variability and wide geographical scope. Mapping the PM2.5 distributions across China with high accuracy and reasonable details is still challenging. Superior satellite-based PM2.5 estimation models need to be developed. Taking advantage of a newly proposed Geographically Neural Network Weighted Regression (GNNWR) model that simultaneously accounts for spatial nonstationarity and complex nonlinearity, we developed a satellite-based GNNWR model to obtain spatially continuous PM2.5 estimates in China. To comprehensively assess the predictive power of the GNNWR model, the widely used Ordinary Linear Regression (OLR) and Geographically Weighted Regression (GWR) models were also carried out for performance comparison. Experimental results demonstrated that the GNNWR model performed considerably better than the OLR and GWR models in terms of multiple statistical indicators, including coefficient of determination (R 2), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Most notably, the fitting accuracy of GNNWR was slightly better than GWR, but its prediction ability was much superior to GWR since the predictive R 2of GWR was significantly improved from 0.683 to 0.831 and the RMSE value was considerably reduced from 9.359 to 6.837. Moreover, the mapped PM2.5 distributions derived from the GNNWR model presented more reasonable and finer details at a higher accuracy than the other models. Although the spatial trends estimated by GWR and GNNWR models were quite consistent, the estimates of the GNNWR model were more accurate and reasonable since its values were much closer to the ground monitoring observations than those of the GWR model, especially for areas with high PM2.5 concentrations, such as Hebei Province and southern Shaanxi Province. In addition, thanks to the excellent learning ability of the neural network, the spatial variations in GNNWR estimates were more sophisticated and displayed a richer hierarchical structure of local changes than that of GWR estimates, which better described the varying details of the PM2.5 across China. In summary, the GNNWR model is a reliable method to effectively estimate PM2.5 concentrations and can also be used to model various air pollution parameters.

Key words: air pollution, PM2.5 estimations, spatial nonstationarity, geographically neural network weighted regression, geographically weighted regression, ordinary linear regression, remote sensing, China