Research On Automatic Extraction Buildings From Remote Sensing Images Based On PU Learning

Buildings play an important role in the production and life of society.The distribution of buildings can help us understand information such as urban development,city functions,population distribution,and provide information support for city planning and policy-making.With the process of remote sensing technology,remote sensing image spatial resolution has a significant improvement in recent decades.It provides the way to extract buildings from remote sensing images.Many internal and foreign scholars are also carrying out researches in this area,and the proposed building extraction methods can be divided into two categories: extraction methods based on physical features of buildings or based on machine learning model.The first method generally used specific algorithms to identify the feature of buildings,and use the feature to extract buildings.It can be run without training samples.However,due to many factors such as complex building shapes,various types of sensors,and different imaging lighting conditions,the building features on different images are unstable,so this method is difficult to use in complex environments;The second method is based on machine learning model,which trains machine learning model through building and nonbuilding labels,and then uses the model to classify unknown label samples.Methods based on machine learning can more effectively gain potential distribution patterns of data,and obtain better results.However,sample acquisition often relies on manually labeling in practical applications.After summarizing the advantages and disadvantages of previous methods,an automatic extraction method for buildings from remote sensing images is proposed,which can work without manually labeling sample.The steps are as follows: First,the image is segmented,and then a method for automatically extracting building samples based on the shape feature and relationship between buildings and their shadows is proposed.A machine learning classifier is trained by the automatically extracting building samples and other unknown label samples.Then use the classifier assign a confidence score to the unknown label samples.Finally,the result is obtained through a dual threshold binarization method based on the graph.The main research contents in this article include are as follows:(1)With the improvement of remote sensing image spatial resolution,images have rich details and redundant information,and traditional pixel oriented image analysis methods are difficult to adapt to this change.Therefore,this paper uses the geographic object based image analysis(GOBIA)method.The basic analysis unit of the method is not single pixel,but a segmented image unit.This transformation can reduce the interference of noise pixels on the one hand,and on the other hand,the segmented object contains more information than pixels,which contribute to target recognition.The foundation of geographical object based image analysis is image segmentation,so this paper selects several common segmentation methods,and analyzes their respective characteristics through experiments.(2)After summarizing previous building extraction methods based on features,this paper propose a building sample extraction method based on shadow presence evidence.This method used the edge features of buildings and the spatial location relationship between buildings and their projected shadows to find strong evidence support of building existence.It can obtain high-precision building samples.Experiment results show that the proposed method have high quality in building samples extracted.(3)A framework for automatic building extraction is proposed.Firstly,shadow features are used to obtain building sample labels,and then positive and unlabeled learning model are trained for extracting building.The method don‘t need manually labeling sample inputs.In order to prove the rationality of the method proposed in this paper,a variety of controlled experiments are designed,and analyze the results based on visual subjective evaluation and multiple evaluation indicators.