Study On Extracting Urban Buildings And Road From Very-High-Resolution Remote Sensing Imagery

In the past 20 years,remote sensing technology has developed rapidly,and highresolution remote sensing image products have made remarkable progress.Remote sensing image data products have accelerated their accumulation and presented the characteristics of mass data and mass storage.In urban areas,80% of ground targets consist of buildings and roads.Buildings and roads are artificial objects that are most closely related to people's lives.For civil use,navigation and positioning,automatic driving,urban planning,disaster prevention and mitigation,and digital city modeling all have greater requirements for the identification of buildings and roads;For military use,target recognition and interpretation,warfare and simulation,and battlefield environmental safeguard.Buildings and roads need to be identified.How to efficiently use high-resolution remote sensing data to quickly extract building and road information has become a hot issue in the field of remote sensing.The traditional object-based building and road recognition methods mainly use the spectral information,geometric information,texture information and semantic information of the target objects.Compared with the pixel-based target recognition method,the robust ability is stronger and the accuracy is higher.In general,the feature set of the object-based method is mainly the two-dimensional feature information extracted from the image objects.It is very difficult to effectively distinguish different target objects of similar.Especially in artificially target objects,the phenomenon called "Difficult objects but same spectrum" and " Difficult spectrum but same objects " is very serious,leading that false detection often occurs.At the same time,with the increasing spatial resolution of images,ultra-high resolution remote sensing image data products,especially sub-meter and centimeter-level,have emerged.The image scenes is more complex,especially in spectral features and texture features.The changes are more abundant,and the phenomenon of "Difficult objects but same spectrum" and " Difficult spectrum but same objects " is more serious.In response to this problem,it is necessary to consider the transition from the traditional two-dimensional information extraction technology to the three-dimensional information extraction technology.In recent years,space satellite technology has matured and airborne platforms have become increasingly abundant.It is easier to obtain satellite stereo or airborne LiDAR data.In particular,the technology of LiDAR combines GPS,Laser ranging,Data mining,and UAV to present all-weather,all-day,fast and efficient features,enabling the acquisition of 3D Information of ground space is easier.In the traditional image segmentation and classification,some problems that are difficult to solve in twodimensional space,three-dimensional data can provide more solutions.However,how to use three-dimensional data in segmentation and classification efficiently is an urgent problem to be studied in OBIA(Object Based Information Analysis).This paper uses the ultra-high resolution(spatial resolution less than 0.09 m)aerial image data and the matching DSM data of a town in Stuttgart,Germany.For the false detection of "buildings" and "roads" in object-based image analysis,based on the classification and multi-level segmentation,a method is proposed to add “threedimensional” features of objects to improve the classification accuracy of “buildings” and “roads” in remote sensing images.In this paper,firstly,combined with the multilevel segmentation and classification,the spectral features,texture features and geometric features of various typical features in the study area are analyzed.By calculating the spectral reflectivity of various objects,an “impervious surface” that does not contain background information such as “vegetation” or “shadow” are mainly extracted.However,because the urban “impermeable surface” is generally constructed by artificial work,and the “two-dimensional” spatial scale is usually very similar,it is very difficult to classify the “impervious surface”.Then,in order to reduce the difficulty of classifying buildings and roads based on the “impervious surface” and improve the classification accuracy of the two class objects,consider adding the stereoscopic feature--“DSM” data.The DSM data is simply organized by the LiDAR points cloud,so the data containing all the three-dimensional “surface points”,which is disadvantageous for the remote sensing image analysis focusing on the “ground objects” extraction,In order to highlight the stereoscopic features of "ground objects" based on object remote sensing image analysis,the strategy of "weakening" or "removing" the influence of surface fluctuations is proposed.The method of object-based stereo layering is proposed to "weak" topographic relief and the method of creating nDSM data is proposed to “remove” the influence of topographic relief.Both methods are based on the object-oriented perspective and the actual needs of the classification in remote sensing image analysis,avoiding the complex LiDAR data processing.At the same time,in order to study the importance of the "stereoscopic features" used in the classification process of buildings and roads,combined with the current popular data mining techniques,the "optimal feature subsets" of buildings and road classifications are explored.Through the combinations of different search algorithms,classifier and evaluators,we can get the different optimal subsets.But all subsets contain "stereo features",which is also fully proved that "stereo features" are crucial in the classification of buildings and roads.Finally,in order to prove the rationality of multiscale segmentation and classification system,and the integration of “stereoscopic features” in buildings and roads classifications,and the necessity of feature space optimization,some experiments were carried out.In this paper,based on the "two-dimensional" multi-level segmentation and classification system,aim at the two types of artificial object objects that are easily confused in the "two-dimensional" space have the obvious feature distinction in the "three-dimensional" space,and fully consider the influence of " topographic relief ".this paper proposed two methods of using DSM data in building and road extraction.Through experiments,the accuracy of the classification model of buildings and roads is 97.95%,96.20%,and the accuracy of models not adding stereo features is 86.08%.The the highest classification accuracy of buildings and roads in test area is 86% after incorporating stereo features,but not incorporate stereo features is only 73%.The classification accuracy is improved by 13%.Experiments show that the proposed method can effectively suppress the phenomenon of "Difficult objects but same spectrum" and " Difficult spectrum but same objects " in the analysis of important artificial target objects;at the same time,it can effectively reduce the dependence of the spectral features and texture features in traditional objects information extraction,improve classification accuracy and meet the requirements of automatic mapping.In addition,it provides a new idea for the two-dimensional information extraction technology to the three-dimensional information extraction technology.