Research On Spatial Air Quality Estimation By Exploiting Terrain Features And Multi-view Learning

With the development of building smart cities,massive amounts of data are generated every day,such as traffic data,air quality data,pedestrian volume data etc.These data come from our lives and will play a vital role with development of the city in the future.The rational and effective use of these data to dig deep into the potential regular meaning of urban data can improve the living environment of the city to a certain extent,with inestimable social and economic value.Air quality data can be obtained from air quality monitoring stations.Since the number of air quality monitoring stations is extremely limited,this puts a demand on air quality spatial estimation.Air quality assessment refers to inferring air quality data for any fine grain area without an air quality monitoring station.At present,the most advanced spatial estimation method for air quality index,in addition to using the air quality index of the surrounding sites,will also consider external factors such as traffic volume,human traffic,POI,etc.,and then build an estimation model based on machine learning techniques.Although these methods have achieved good results in urban areas,it is difficult to ensure the effects in non-urban areas(such as waters,mountains,forests,etc.),this is because:(1)The existing method does not consider the influence of topographic factors on air quality,and uniformly models the air quality spatial estimation problem in all terrain types,resulting in the model's generalization ability in special terrain types is not strong.(2)There are too few air quality monitoring stations and most of them are deployed in urban areas,resulting in too few training samples(especially in non-urban areas).To solve these problems,this paper intends to start from the air quality space estimation project,focusing on the air quality space estimation model which can have strong generalization ability in various terrain types,and then expound the method of mining urban data.The paper mainly carried out the following work:(1)Terrain feature mining.At present,most existing urban air quality spatial estimation methods mainly focus on urban areas to extract features,while good results are usually not achieved in non-urban areas.Therefore,in order to accurately estimate the air quality in non-urban areas,this paper builds a terrain database based on open source maps,then mines the terrain feature,including extracting the terrain type distribution at local and nearby stations,and refining the original terrain features based on the integrated decision tree model.(2)Multi-view transfer semi-supervised learning.The number of air quality monitoring stations is very limited,and most of them are deployed in urban areas.In most cities,there are even no air quality monitoring stations deployed in special terrain.Therefore,for a particular city,only relying on data generated by the city's air quality monitoring station cannot effectively establish a spatial assessment model for air quality.To this end,this paper proposes a multi-view migration semi-supervised learning model(MTS4AE),which comprehensively utilizes the annotated data generated by the air quality monitoring stations in other cities with similar special terrains and the unlabeled data of the city's airless quality monitoring station area to establish a spatial estimation model for air quality.First,the features are divided into three categories: time series features,spatial features and deep features.Then,based on the labeled data generated by the air quality monitoring station in similar terrains of other cities,the migration learning algorithm is used to establish an initial model for each type of feature.Finally,based on the unlabeled data of the city's airless quality monitoring station area,the collaborative learning algorithm is used to fuse the initial model.