A Geometric Saliency For Building Extraction In High Resolution Remote Sensing Images

As a basic geographical element,building and its accurate map are widely used in urban planning,navigation,disaster emergency and so on.With the development of remote sensing,we can now easily obtain high-resolution remote sensing images which including lots of buildings.High resolution image has the advantages of fast update speed and the boundary of building is accurate,which can meet the urgent needs of various applications for accurate building distribution map.Therefore,how to extract buildings from high-resolution images has become an important and challenging task.However,some existing building extraction methods focus more on the detection of building area,ignoring the preservation of the geometric contour of a single building.In this paper,a new method of building extraction is proposed to keep the accurate geometry of buildings in high resolution images.By analyzing the geometrical features in high resolution images,we observed that buildings have higher distinguishability in geometric than in texture or spectral feature space.At the same time,geometric features naturally have the characteristics of describing the accurate boundary of buildings.Based on this characteristic and the further experimental analysis of the unique distribution of geometric features in the building area,we model the relationship between geometric features and buildings and construct a building index calculation framework based on junction,which can effectively distinguish the buildings with natural objects,and meanwhile maintain the geometric shape of each single building.We compared our algorithm with six different methods and the experimental results on three public datasets show that the proposed algorithm has high building detection accuracy,and can distinguish the exact location and shape of a single building,which benefits from the accurate description ability of geometric features to the building boundary.It is worth mentioning that the proposed algorithm also has good generalization performance,and can still show good detection performance in the case of large distribution differences of buildings.In addition,through the analysis and experiment of different levels of geometric feature information,we find that the angle information is helpful to distinguish the buildings from the background area.This is mainly because the building has the characteristics of shape regularity,and there is angle difference between the building and the background features in the geometric feature space.At the same time,by replacing the feature extraction module algorithm,we further prove that the overall algorithm framework has good stability on junction detection algorithms.