Automatic Building Extraction From Satellite High Spatial Resolution Imagery Using Mathematical Morphology

Buildings are one of the most important types of artificial targets in the urban environment.Due to the high frequency of changes in buildings,understanding their current distribution is important for urban planning,change detection,urban environmental investigations and urban monitoring applications.The use of a new generation of very high spatial resolution sensors,such as Ikonos,Quick Bird and Worldview,has broadened the application of remote sensing technology.A great amount of spatial and thematic information on land cover at local and national scales is contained in satellite very high-resolution(VHR)remote sensing images,and this information gives buildings clearly identifiable shape and texture features.In view of this,VHR images are suitable for building extraction tasks.However,the high intra-class variance as well as the low inter-class variances in the spectral statistics of VHR images greatly reduces the distinguishing ability of small land-cover areas in these images.To address this problem,numerous studies have focused on the extraction of spatial and structural information in images and the use of this information as a supplement to improve the recognition ability.Researchers have indicated that importing spatial features significantly improves the accuracy of VHR image classification.Morphology attribute profiles(APs)provide a flexible way to model information from high-resolution images,the transformations in APs can extract features based on either the geometrical or spectral characteristics of objects.It have been wildly used in the processing of remote sensing image,and have proven to be effective tools for spatial feature extraction tasks.However,instead of acting as an automatic image-processing index,APs often work as ancillary features of the spectral characteristics in supervised learning.That is,the intrinsic land-cover recognition ability of APs may be underestimated,prompting researchers to continue to study it.The target of this paper is to automatically extracting buildings in the high-resolution image based on morphological operators.To obtain the target of extracting buildings automatically and accurately in the image of dense urban area,the research is carried out based on non-building feature extraction,building feature extraction and shadow feature extraction,and the proposed method is verified by the real urban remote sensing datasets.At frist,we analysed the effects and performance of different supervised machine-learning approaches for building detection tasks.After expounding the spectral and geometric features related to building extraction in high-resolution images,two representative supervised classifiers: support vector machine and random forest.Both pixel and oriented based strategy are used to test the building extraction performance on the experimental datasets.The experimental results show that the object-oriented method obtain more precise than the pixel-based method after importing the geometric features of the object.However,there are still high error rates in their extraction results.To obtained a better building detection precision,the supervised classification is complicated and time-consuming in manual representative samples selection stage,which cannot meet the automatic processing requirements of large-scale data.In recent years,there are many studies focus on the automatic extraction of buildings.Different strategies,such as a number of building feature indexes have been proposed to extract potential buildings or exclude confused non-building features,such as vegetation,water and shadows,the accuracy of building detection can be improved by eliminate these background.In the existing building feature indexes,the morphological building index(MBI)have been proven to be effective tools for automatically detect buildings in VHR images.By analyzing the MBI algorithm,it is found that MBI is subject to the high omission and commission rate.The reasons for the high omission rate are: the effect of the dark details contained in the building roofs;some missing dark roofs.The high commission rate is derived from the false alarms of the landcovers that have similar spectral characteristics with building.To detect the buildings in urban high-resolution image,this study is focuses on the automatic non-building feature extraction,automatic building feature extraction and automatic shadow feature extraction.Firstly,for the problem of dark details in the building roofs and misdetection landcovers,two strategies are proposed: 1)image denoising method based on morphological attribute filters,and 2)elongated non-building object extraction method,to deal with the two problems,respectively.The former one is used to filters out the dark details while improve the interior homogeneity of the bright connect componen by measures the spectral homogeneity of each regions.By analyzing the different shape characteristics between buildings and non-buildings,the latter strategy uses attribute morphological filtering to separate the buildings from the non-building objects that with similar spectral characteristics,so that the buildings remain independent and complete outline.The experiment results show that the proposed methods can effectively improve the integrity and accuracy of building extraction result and reduce the omission and commission errors in the results.Secondly,a new building feature index based on morphological attribute profile is proposed,which is called morphological attribute building index(MABI).By investigating the associated attributes in morphological attribute filters,the proposed method establishes a relationship between AFs and the characteristics of buildings in high resolution images(e.g.,high local contrast,internal homogeneity,shape and size).By the sequential application of attribute filters,multilevel characterization of the image is obtained to model the structural information of buildings.Considering the different reflectance characteristics of buildings in the image,the bright buildings and dark buildings are extracted separately in the MABI to reduce the omission rate caused by the absence of dark roofs.By analyzing the building extraction results with existing methods,it is show that the proposed one obtain the most accurate result combined with high speed and high degree of automation.The proposed algorithm does not need any training samples,and it is suitable for processing large-scale data.Finally,in order to further reduce the ommision and commission errors cuased by the single threshold segmentation of the MABI,a building extraction framework is proposed.This framework consists of two main parts: 1)shadow extraction and and 2)dual threshold object-oriented analysis.A new automatic shadow extraction index,called morphological attribute shadow index(MASI),is proposed in 1).As one companion of building,the spectral and geometrical characteristics of shadows are opposite and similar,respectively,to the corresponding characteristics of adjacent buildings.2)A dual threshold analysis strategy.This strategy is aimed at the problem of high commission rate caused by the use of a single threshold for MABI feature segmentation.Under the object-oriented framework,different spatial constraint relationship between the shadow and the building for objects with different MABI values is considered to suppress non-building objects.Through experimental comparison,it is found that this building extraction framework can effectively improve the detection accuracy.The image denoising,non-building object extraction,building and shadow extraction methods proposed in this paper can automaticly extract buildings from high-resolution images without any prior knowledge and supplementary information.The proposed framework obtain high accuracy and algorithm efficiency in the building extraction experiments.Compared with the existing supervised and unsupervised algorithms,the proposed strategy has achieved the optimal extraction accuracy..