Multi-view Subspace Clustering Methods For Inferring Urban Land Use From Multi-source Geospatial Big Data

In the era of big data,geospatial big data brings new opportunities for inferring urban land use from the perspective of social function,which is of great value for promoting people-oriented urban planning and governence.Spatial clustering is the main technique for inferring urban land use types from geospatial big data.Most of existing works based on spatial clustering methods infer land use types from single type of geospatial big data.Due to the bias of information from single source geospatial big data,the inference results will not be enough to depict the comprehensive view of urban land use.At present,challenges remain in fusing multi-source,high-dimensional and noisy geospatial big data to infer urban land use.In this thesis,multi-source geospatial big data was regarded as different views to reflect urban land use.Multi-view subspace clustering methods were employed to fuse multi-source geospatial big data to infer urban land use types.The main contributions are as follows:(1)Comparison and evaluation of urban land use inference methods based on spatial clustering analysis.By using taxi trajectory data and bus smart card data of Beijing in 2016,existing urban land use inference methods by using a single type of geospatial big data and multi-source geospatial big data were compared and evaluated.The results showed that:(1)the accuracy of land use inference by using a single type of geospatial big is low due to the bias of data;(2)The performance of the weighted fusion method is better than that of the method by using a single type of data,but the detection rate of land use types is still low(the overall accuracy is44.54%),and it is difficult to define the weights of different data sources.(2)Urban land use inference based on a latent multi-view subspace clustering method.To handle the biased,high-dimensional and noisy geospatial big data,a latent multi-view subspace clustering method was used to infer urban land use types.Firstly,the variation in the number of origion/destination points over time of multi-source geospatial big data was used to characterize land use types;Then,a latent multi-view representation was applied to construct the common structure shared by multi-source geospatial big data;Finally,based on the latent multi-view representation,the subspace clustering method was used to infer land use types.Experiments on taxi GPS trajectory data and bus smart card data in Beijing revealed that,compared with the method using a single type of geospatial big data and the weighted fusion method using multi-source geospatial big data,the latent multi-view subspace clustering method can obtain the highest detection rate of land use types(the overall accuracy is57.77%),and avoid artificially setting the weights of different data sources.(3)Urban land use inference based on an adaptive graph constrained multi-view subspace clustering method.Most existing latent multi-view subspace clustering methods only impose constraints on the data reconstruction error,which results in a risk that the neighboring relationships of data points is difficult to preserve in the subspace.To overcome this limitation,an adaptive graph constrained multi-view subspace clustering method was developed for inferring urban land use types.Firstly,the autoencoder networks were used to obtain the latent representation for each type of data;Then,in the framework of subspace clustering,multiple latent representations were fused into a shared subspace representation,and a shared nearest neighbor graph was used to constrain the local structure of original feature space and subspace.Finally,the subspace representation was used for spectral clustering to obtain the clustering results.Experiments on four multi-view benchmark datasets showed that proposed method outperforms nine representative methods.Experiments on taxi GPS trajectory data,bus smart card data and points of interest data of Beijing in 2016 showed that compared with the method using a single type of geospatial big data and the weighted fusion method and latent multi-view subspace clustering method,the proposed method achieved the highest land use detection rate(the overall accuracy is68.11%).The clustering results can provide reference for urban planning.