基于半监督小样本学习的北京市臭氧空间分布模式分类研究  

Classification for Spatial Patterns of Urban Ozone Pollution in Beijing Based on Semi-Supervised Few-Shot Learning

作　　者：孙津 SUN Jin(School of Earth Sciences and Engineering,Hohai University,Nanjing 211100,China;Jiangsu Province Engineering Research Center of Water Resources and Environment Assessment using Remote Sensing,Hohai University,Nanjing 211100,China)

机构地区：[1]河海大学地球科学与工程学院,南京211100 [2]河海大学江苏省水资源环境遥感监测评估工程研究中心,南京211100

出　　处：《地球信息科学学报》2024年第3期725-735,共11页Journal of Geo-information Science

基　　金：中央高校基本科研业务费专项资金项目(423062)。

摘　　要：研究城市臭氧空间分布模式有助于分析污染成因,也能够为污染防治提供科学依据。但过去多基于站点数据分析浓度平均分布的概括性特征,对全域浓度分布平面的描述有限,也很少进行分类研究,无法更全面地看待污染分布的多种模式及其时间变化。本研究利用由卫星数据建模估计的臭氧日最大8 h滑动均值分布数据,针对臭氧空间分布模式标签化的难度提出一种面向小样本的半监督学习方法,以北京市为例进行分类实验。实验发现:(1) 2020年数据经预处理后以40个训练样本对249个测试样本进行分类,总体分类精度达81.12%,kappa系数达0.741 6,说明在小样本条件下半监督方法取得了较好的分类效果;(2)分类得到的8种模式中,“(东)南高(西)北低或东高西低”的模式1、“(西)北高(东)南低”的模式2以及“中心低”的模式6为主要模式,分别在暖季(3—10月)、冷季(11—次年2月)和冷暖季过渡期占据主导,这一时间规律反映出区域传输和光化学反应的季节性影响;(3)将2020年的训练样本迁移至2019年进行分类,在取得较高精度的同时也对上述规律进行了验证。以上结果表明,本研究提出的空间分布模式分类方法能够为全面确定高污染的防治区域以及分类研究不同污染事件的成因提供支持。Ozone concentrations tend to be heterogeneous across a city's space due to the mixed land use and diverse landscapes.Studying spatial patterns of urban ozone pollution contributes to the knowledge of the mechanism of pollution formation and also provides scientific reference for pollution prevention and control.Nevertheless,most previous research focused on the averaged value of ozone concentration from monitoring sites,which cannot describe the spatial characteristics of the entire region's concentration surface.Additionally,the classification method was seldom used to analyze pollutants'spatial patterns,and thus very few studies paid attention to the varied types of patterns and their temporal variations.In this study,based on the distributions of ozone’s daily maximum 8-h moving average estimated from satellite data,an approach of semi-supervised few-shot learning was proposed to classify ozone's spatial patterns in Beijing.The self-training method considered the difficulty of data labeling and can utilize information from a large number of unlabeled samples to augment the training set iteratively.Three kinds of normalized features were involved in classification to describe the spatial variations of concentrations.Totally,there were 40 training samples and 249 test samples for the year of 2020,and the overall classification accuracy was 81.12%with a kappa coefficient of 0.7416.This demonstrated the effectiveness of the semi-supervised classification method despite the small size of training samples.The classification results showed that,among the eight patterns of ozone distributions in Beijing,three of them were major patterns,including Pattern 1:high concentrations in the south/east/southeast and low in the north/west/northwest,Pattern 2:high concentrations in the north/northwest and low in the south/southeast,and Pattern 6:low concentrations in the center.They dominated the warm season(from Mar.to Oct.),the cold season(from Nov.to Feb.),and the transition period,respectively.These temporal variations of

关 键 词：臭氧 空间格局 半监督分类 图像分类 小样本学习 样本迁移 季节性 北京 

分 类 号：X515[环境科学与工程—环境工程]