Research On Object-Oriented Urban Land Use Classification Based On Social Media Geographic Data

For many fields of research,the urban land use data is indispensable.How to obtain high-precision land use data has always been a difficult and hot topic in the relating field,which is very practical.Object-oriented classification methods have been widely used in detailed land use classifications,by which remarkable results were obtained.However,the existing methods still relied on the interpretation of high-resolution remote sensing images,leading to great limits in practical applications.The development of social networking services,spatial information services and communication technologies has spawned massive amounts of social media geographic data.As an important part of the massive geographic information era,social media geographic data shows incomparable advantages while compared with traditional data sources:rich spatiotemporal and textual information,large volume,fast updated,closely related to crowds' activities,open source,easy to be acquired,etc.By utilizing social media geographic data,the activity trajectories generated by great amount of users can reappear.Generally,people undertake different activities in urban zones of various land use types.By mining the spatiotemporal and textual information in social media geographic data,the spatiotemporal patterns and topics of crowd activities can be understood and applied in the object-oriented land use classification,which could take place of the high-resolution remote sensing images.Thus by using this kind of data,the high accuracies in results can be guaranteed.Based on full investigation of existing researches,we proposed an objected-oriented methodology in which the object was determined as land parcel,and the land use types were distinguished by analyzing the social media geographic data within land parcels.Therefore,we selected the typical Twitter data for an example analysis:relying on Geographic Information System(GIS),with land parcel data and Twitter data,the spatial analysis,data statistics and text mining methods were combined to extract the propertied of parcels,as well as the spatio-temporal and topic attributes of the crowds' activities within each parcel from the timestamp information in Twitter and the Tweets(textual part in Twitter).According to the obtained multi-dimensional attributes,a supervised learning model was constructed for land use classification,and the overall precision of results reached 83.8%finally,proving the effectiveness of the method.The specific contents of our research are as follows:(1)Data preparation:taking the geo-tagged Twitter data as example,we used the streamR package in R to capture the data within study area,next performed data cleaning and preliminary pre-processing.In addition,we acquired land parcel data in this area;(2)Attributes acquisition:1.with combination of various spatial analysis methods,the attributes of land parcels were mined;2.according to the characteristics of social media geographic data,we further pre-processed the timestamp information and defined temporal attributes of the crowds' activities.By some Python programs to link the MySQL database,the temporal attributes of the crowds' activities within parcels were extracted;3.Based on the Tweets within parcels,we used the Latent Dirichlet Allocation(LDA)topic model to extract the activity topics of crowds in each parcel respectively;(3)According to the multi-dimensional characteristics obtained,we constructed a neural network model,and applied the Grid Search method to tune its parameters to obtain the best classification results.In addition,different combinations of attributes were set up for comparative experiments to assess their performances in the classification.We further explored the causes for the generations of the classification results.