| Remote sensing technology,which can provide stable data support for urban land monitoring,is considered as the most powerful means to quickly and accurately obtain urban land cover/land use(LCLU)information.Thanks to the rapid development of machine learning,remote sensing image information extraction has transformed from traditional manual visual interpretation to computer automation.However,the acquisition of training samples is a time-consuming and laborious process,and remote sensing images are more suitable for expressing land cover information and have poor recognition effect on land use.The emergence of crowdsourced geographic data provides a new solution for the traditional research on urban LCLU.On one hand,Open Street Map(OSM)has the characteristics of high production speed,short cycle,and has ground object types,which can partially or completely replace the in-field sampling work of professionals,greatly reducing the cost of human and material resources.Traditional characteristics expression of urban land use,on the other hand,is based on the spectral or spectral derivative features such as shape and texture,but land use is closely related to human activities.Social sensing data can provide the space-temporal law of urban residents' activities.This humanities characteristics can effectively solve the limitations of urban land use extraction by only using the natural features extracted from remote sensing images.This paper combines high-resolution remote sensing images and two typical crowdsourced geographic data--OSM data and social sensing data,to develop the LCLU research in urban areas successively.The research builds a complete urban LCLU research framework in terms of the land cover sample construction based on the OSM data,the multi-features extraction of land use and the elaborate classification of urban land use,respectively.The approach is conducted in the central area of Wuhan to discuss the application of combining professional remote sensing data and nonprofessional crowdsourced geographic data in the study of urban planning and management.Contents of the research in this paper are as follows:(1)In this paper,a land cover sample construction method based on OSM data is proposed.The corresponding relationship between OSM category and the required land cover category is established to extract relevant sample labels from OSM.Morphological errosion is utilized to eliminate the spatial errors caused by the mismatch of the feature boundaries.And a threshold selection method of remote sensing index is adopted to remove the attribute errors of OSM data caused by wrong marked objects.As a result,the training samples extracted from OSM are purified.(2)Since the structure of ground objects in high-resolution remote sensing images is clear,spatial features extracted from the spatial relationship of pixels can improve the separability of different land cover types in the feature space.Based on the spatial structure characteristics of high-resolution remote sensing images,the combined use of spectral and morphological attribute profile features solves the serious mixing problems when only using spectral features.(3)The parcels generated by OSM road network is used as the basic research unit of land use,and a multi-feature system for identifying urban land use information is established from three aspects of physics,landscape and humanity.Firstly,spectral features and texture features of the parcels are extracted from high-resolution remote sensing images to form physical features.Secondly,on the basis of the land cover map generated in research(2),landscape index is introduced to describe the spatial heterogeneity of land cover composition and structure in land use units.The landscape features of land use units are described from class scale and landscape scale.Finally,social sensing data,which can reflect the time and space change of different public groups activities,are exploited to reflect the humanity features of land use.This paper studies the adjacent time changing trends and shapelets features based on the subsequences to form comprehensively cultural characteristics that describe the population change in parcels.(4)The ability to identify urban land use types by simply using remote sensing image or crowdsourced geographic data is limited,leading to a rough set in traditional land use classification research.This paper established a fine-grained land use classification system by using multi-features to jointly describe the characteristics of land use types.Aiming at the characteristic of imbalanced high-dimensional small samples,the method of sample balancing is discussed.Also,support vector machine(SVM)and random forest(RF)which are suitable for training high dimensional small samples are selected as classifier.The feature importance evaluation and the uncertainty estimation of classification results is proposed for analysis.(5)In this paper,the area within the third ring road of Wuhan(the central activity area)is taken as the experimental area.The GF-2 remote sensing image data,OSM data and Tencent user location big data were collected.The land cover classification and land use classification of central activity area of Wuhan were successively conducted.In the land cover research,the proposed method is compared with classification results under the conditions of using unpurified OSM original samples,using no spatial features,using roads as training samples instead of superposition strategy and using object-oriented methods,etc.,and is compared with a variety of other existing methods.In the research of land use in Wuhan,aiming at the overfitting problem of high dimensional small sample,the comparison experiments of different sample balancing methods were conducted.The impact of different combinations of physical,landscape and humanistic characteristics on classification results are explored.The importance of different features and the uncertainty of the classification results was evaluated.Finally,a comprehensive analysis of the spatial distribution of land use in Wuhan was conducted. |
PDF Full Text Request