Study Of Urban Village Detection Methods Based On High-resolution Remote Sensing Imagery

In recent decades,China has experienced a rapid urbanization.As a result,urban villages(UVs)are common in many cities,which become a challenge to urban man-agement and planning.Despite wide attraction of UVs for researchers,current studies on UVs are severely incomplete,e.g.lacking of quantitative spatiotemporal analysis,etc.In the remote sensing community,it is widely acknowledged that high-resolution remotely sensed imagery has become an important data source for urban planning and management,however,there are almost no remote sensing related studies on UVs.It is urgent to use high-resolution remotely sensed images for extraction and analysis of UVs,which benefits urban managers and planners as well as researchers from other fields.Currently,middle-level feature based,neural network based,and object-based methods are three primary kinds of methods for scene extraction from high-resolution remotely sensed images.As far as UV extraction,however,there are some problems in these methods:(1)middle-level features cannot effectively describe UV semantics;(2)neural networks have powerful ability for representing scenes,but they need lots of data,especially labeled data,which costs much labor and time;(3)image segmentation cannot reliably produce accurate objects due to the complex texture of UV scenes,and there is no mature approach to representing relationships between various objects.To this aim,this paper proposes two UV extraction methods respectively based on deep neural works and scene semantic description,and establishes a UV monitoring framework for multitemporal and multi-regional high-resolution remotely sensed images.The paper also quantitatively analyzes the spatiotemporal patterns of UVs of Shenzhen and Wuhan.Details of the paper are summarized below:1.Based on deep neural networks,an unsupervised deep feature learning method is proposed for UV extraction.The proposed method consists of an unsupervised deep convolutional neural network(UDCNN)and an unsupervised full-connected neural network(UDFNN).UDCNN first performs multiscale and adaptive feature extraction on scenes,and then feature pooling is used for dimension reduction.Finally,UDFNN is used to extract the internal structures of the features,resulting in deep abstract scene representation.2.A UV extraction method based on scene semantic description is proposed.The methods exploits indexes for the fast extraction of semantic objects(i.e.buildings,vegetation),and then describes the proportion and composition of these objects in the scene by two proposed scene semantic descriptors.The first descriptor constructs scene semantic features by histograms of object attributes,and the second descriptor is based on landscape metrics,which are used to represent spatial relationships between objects and the scene.In the experiments,the deep neural network method and the scene semantic description method outperformed other unsupervised feature learning methods and traditional middle-level feature based methods,while the two proposed methods achieved comparable results.3.A UV monitoring framework is established for multitemporal and multi-regional high-resolution remotely sensed images.The framework extends the scene semantic descriptor based on landscape metrics,and it can use labeled samples from other images.Transfer learning is introduced to address the inconsistent distribution of sample features from different images,and a re-selection strategy is proposed for transferring landscape metrics across different study areas.4.The proposed methods are applied to high-resolution remotely sensed images of Shenzhen and Wuhan,and,based on the extraction results,quantitative spatiotem-poral change analysis is performed.The City-UV-Building hierarchical structure is analyzed using landscape metrics.Multitemporal changes of UVs at the city scale and their statistical distribution at the UV scale are presented.By spatial clustering,a significant correlation between the development and the location of UVs in Shenzhen is revealed.Finally,two strategies against the redevelopment of UVs are discussed and summarized.