Research On Object Recognition Method Of Urban Buildings In High Spatial Resolution Remote Sensing Imagery

High spatial resolution remote sensing image can accurately express rich surface information.It is a research hotspot in the field of high-resolution remote sensing image analysis and application to recognize urban buildings by using the macro and full coverage of high-resolution remote sensing image.However,the current research still lacks the effective technical methods to quickly and accurately transform the surface elements from the remote sensing image space to the geographic information space.The existing methods or algorithms still have obvious limitations when they are applied to the object-oriented recognition of high-resolution remote sensing urban buildings:(1)The recognition tasks of different elements have certain pertinence to feature selection and need rich experience and knowledge,which leads to the lack of universal application of the method.(2)The object-oriented recognition of image targets in high-resolution images is mainly realized by multi-scale segmentation technology,but it is still not fully competent for the task of automatic information extraction of complex ground objects(such as buildings).At present,there is no systematic and efficient method and technical system to solve the problem.The rapid development of deep learning technology in the field of computer vision has brought new opportunities for the object-oriented recognition of thematic features based on high-resolution remote sensing images,but there are still two dilemmas:(1)There is still a lack of high-precision building contour level sample database constructed by multi-scale hierarchical logic,and the type organization mode,scale distribution characteristics and sample data quality of the sample data will have a great impact on the application effect of deep learning;(2)At present,the mainstream method in the world adopts the example(or semantic)segmentation method of pixel level mask.The output results of the model can not be directly applied to the production of GIS vector data and the data quality can not meet the relevant data quality standards.The real case segmentation for building vector polygon extraction needs to be carried out urgently.Based on the advantages of deep learning in image target recognition,and based on the geometric spectral characteristics of building elements in high-resolution remote sensing image,this research is aimed at the engineering application of object-oriented recognition method for high-resolution remote sensing big data buildings,and has carried out the research work of vector level target extraction of high-resolution remote sensing urban buildings.The main research contents and application results of this paper are as follows:(1)Image preprocessing and multi-scale target sample construction for high-resolution remote sensing urban building object recognitionIn this study,the visual features of high-resolution images are enhanced by image preprocessing,and the comprehensive building semantic features and multi-scale spatial spectral features are input into the network model to further optimize the process of spatial-temporal feature convolution.Considering the bit adaptation of the deep learning network and the enhancement of the spatial spectral characteristics of the image block,we have processed the linear stretching and image filtering of the constructed area to be recognized,which not only suppressed the noise in the image data,but also preserved the detailed feature information of the image data as completely as possible,thus enhancing the image interpretation and recognition effect.On this basis,we extract multi-scale semantic target samples of urban buildings for deep learning,effectively complete the construction of building case samples and generate high-quality image data sets,and form a complete set of technical system for building deep learning sample data sets,which meets the requirements of object-oriented recognition of urban buildings for high-resolution remote sensing images in this study And the need of building vector extraction.(2)Object recognition of high-resolution remote sensing urban buildings based on Mask R-CNNIn this paper,we transfer the training feature weights of coco data set to Mask R-CNN,and integrate the migration learning mechanism of sample database,so that the network learning can be self-organized and self-adaptive enhanced,and then combine the spacespectrum features of high-resolution remote sensing training image,the relationship between adjacent pixels and other feature factors to participate in the final decision classification.In the follow-up experiments,we carried out super parameter experiments and structural modifications of the middle and high-level network.At the level of network architecture,we focused on enhancing the extraction of features of the middle and high-level,and constructed a corresponding deep neural network to study the special features.It broke through the limitation of the expression and calculation ability of the Traditional high dimensional features,and realized the objectification of urban buildings with high-resolution remote sensing image Recognition is instance segmentation,which lays a system and data foundation for the following empirical research on the application of urban building recognition oriented to building vector polygon extraction.(3)The application of building recognition based on building vector polygon extractionOn the basis of(1)and(2)research contents,we use high-resolution remote sensing satellite(aerial photo)data series to carry out the implementation scheme design and empirical research of urban building object recognition in Mandalay City,Myanmar(the buildings are scattered in different levels,and the samples are representative).Empirical work studies and establishes a set of object-oriented recognition technology system of high-resolution remote sensing big data buildings for building vectorization polygon extraction,which verifies the above-mentioned object-oriented recognition method of high-resolution remote sensing image urban buildings.The experimental results show that the technology flow system of this paper can meet the needs of engineering practice in different experimental areas,and there are obvious differences in spatial spectral characteristics of image data.Based on the feature information of building elements of high-resolution remote sensing image,and based on the engineering application of object-oriented recognition method of high-resolution remote sensing big data,this paper studies the extraction of target level and vector level of high-resolution remote sensing urban buildings.The empirical results show that the method flow proposed in this paper can be applied to the effective extraction of city level buildings,expand the theoretical basis of the object-oriented recognition method of remote sensing thematic features based on the idea of object image analysis,and basically solve the problem of building information extraction and recognition in high-resolution remote sensing.