Reserch On Building Extraction From High-resolution Visible Optical Remote Sensing Images

Buildings are the main man-made objects that constitute the living environment of human beings,and the most critical urban elements for the future smart city.In order to achieve effective support for smart cities,accurate and reliable building extraction is the most critical information to achieve the 3D building targets in city block/community.Therefore,with the fast development of remote sensing technique,a lot of researches for building location and boundary extraction have been carried out all over the world in the literature.However,although “experience” or “knowledge” of various building inspections have been formed,they are limited by the ability of the building detection process to deal with problems caused such as noise,occlusion,and illumination in the image.The processing power to extract the original clues required to reveal the intinsic building prior knowledge or rules has not been better guided by relevant knowledge.Obviously,accurate building extraction is still the hotspot and problem in the literature.Specifically,in order to realize accurate and reliable building extraction from high-resolution visible optical remote sensing images,it is necessary to effectively solve the following key problems:(1)How to effectively realize the comprehensive application of knowledge such as spectral information(that is,color information),multi-scale structure information,spatial combination rule information in visible light remote sensing images,it is need to form an effective macro strategy level.(2)How to efficiently use the available spatial information and structural information to reduce the impact of noise,occlusion,lighting condition and other issues,to accurately and reliably extract building edges and complete contour information to give support for the overall building extraction system.(3)How to fully combine the multi-scale region and contour information contained in the visible opitcal remote sensing images to achieve accurate and reliable building targets extraction.In order to address the above problems,typical process of current building extraction technology and the main building extraction rules were analyzed in detail in this dissertation.Guided by the idea of mining building targets related information in an active way,research work on the following aspects has been carried out:(1)Aiming at the problem that the edge extraction techniques used in the existing building extraction methods are always sensitive to noise and interference(roads,tall trees,regular vegetation,etc.),an building edge extraction method is proposed in the third chapter based on the joint of orientation and color information with the understanding the macroscopic characteristics of human information perception system.The proposed method can significantly improving the performance of building edge extraction:Most of the existing building extraction methods or systems use Canny,Sobel or other common edge extraction methods to acquire the initial edge map.It means that,in the preprocessing stage,the building target features or rules are not fully utilized,which directly affects the accuracy of subsequent building target extraction tasks.By analyzing the characteristics of human perception procedure,with the effective mining of color information and direction information as the core,the edge sensitive color components and the main direction information of the buildings are extracted effectively.And then,with these building priors,an enhanced dual-window Gaussian gamma shaped filtering(GGS)method is proposed to extract edge maps for building extraction,which provides a good foundation for obtaining accurate and complete contours of building targets.(2)Aiming at the problem of weak constrast of building edges in high-resolution remote sensing images caused by lighting and material issues,which will caused incomplete edge information given by the initial building edge extraction.From a classification point of view,combining dual-scale Sparse-SVM edge classification and decision fusion strategy,a novel building coutour recognition and extraction method is proposed in the fourth chapter to improves the integrity and reliability of building object extraction:In the edge map achieved by building edge extraction from the visiable optical high-resolution remote sensing images,there will inevitably be false edges caused by disturbances such as roads and vegetation,weak edges and leaky edges due to the color of the building roof and the weak contrast area.In order to better extract the contour information of the building target in this case,a building edge optimization and grouping method combining pixel-level edge classification strategy is proposed to realize the accurate and reliable extraction of the building target contour.(3)Aiming at the problem of insufficient fusion of contour information and regional information in existing building object extraction methods,a multi-scale deep learning method for extracting salient area of building objects is proposed in the fifth chapter,which effectively improves the accuracy and reliability of building object extraction:The building countours extracted from the high-resolution visible optical remote sensing images are often irregular,and the corresponding edge positions are always affacted by noise and other interferences.In order to achieve the objective of accurate and reliable building extraction,regional information or salience regional information will be a certain beneficial compensation for the final building extraction.Therefore,considering that existing deep learning building extraction methods do not make full use of multi-scale semantic information,with the fully convolutional neural network(FCN)as base machine,a multi-scale deep learning based saliency regions extraction method is proposed to realize the efficiently combination of contour and regional information of condidate building targets.Finally,accurate and reliable building targets can be extracted.In order to verify the effectiveness of the proposed method,multiple satellite and aerial high-resolution visible light remote sensing datasets related to the construction of 3D cities and smart cities were used for conducting comparison experiments.The experimental results show that for the building object extraction method proposed in this paper,the utilization logic of the relevant information is reasonable,and it is sufficiently robust to the target areas of different buildings with different density and different main directions,and has good application prospects.