Research On Man-made Target Detection Of High-spatial Resolution Remote Sensing Imagery Based On Multi-scale Morphological Features

The rapid development of earth observation technology makes it easier to obtain highspatial-resolution remote sensing images(HSR),and how to gain and make use of the information from these massive HSR image data,accurately and efficiently,is still a problem that needs to be solved.The accurate location and area information of man-made targets represented by buildings and roads not only plays an important role in urban planning,socioeconomic environment analysis,traffic management and other fields,but also critical for building basic geographic database and smart city construction.In recent years,great progress has been made in buildings and roads detection from HSR images,but there is still a gap between its performance and practical applications.The difficulties of building and road extraction from HSR imagery mainly consist of followings: 1)The HSR imagery is actually amplified the scale of observations of Earth’s surface,and the highly detailed information lead to the spectral heterogeneity increased within a class and are decrease between classes.Moreover,the phenomenon of “same object with different spectra” and “different objects with same spectrum” is also remarkably.2)The multiresolution characteristics of HSR images makes it difficult to fully use of multi-scale information from images based on single-scale analysis methods,and therefore cannot obtain the optimal outputs.3)The semantic environment of HSR remote sensing images is complex.For example,The spatial topological discontinuity and high regional spectral heterogeneity of roads caused by shadow and vehicle occlusion in HSR images degrade the robustness of existing methods.Those problems pose great challenges for the application of HSR images for building and road extraction.This research focuses on the multi-scale morphological feature analysis of HSR image and its application in building and road detection.The key techniques and solutions in the extraction of HSR images of buildings and roads were studied and proposed.The specific research content and results are as follows:(1)An unsupervised building extraction approach based on a combination of probabilistic model and morphological attribute filtering is proposed for non-urban areas(mainly in densely populated areas such as relatively small towns and rural areas).The method first extracts the residential areas by probability density estimation.Then,in order to accurately determine the location and area of buildings,the multi-scale morphological attribute filtering is applied to the image and the corresponding morphological attribute building indexes are constructed to complete the building extraction.Finally,the morphological filtering is applied to the image with the attribute named "moment of inertia" to extract the road networks in residential areas,and a semantic constraint criterion is built to refined the building extraction result.The quantitative numerical comparison experiments verify the effectiveness of the proposed method in building extraction in non-urban areas.(2)The core idea of the construction of the GMABI is that the scale-span differential operation of attribute profiles is used to substitute for the adjacent-scales differential operation which used in the construction of the morphological building index and morphological attribute building index,therefore,obtained more complete image structure spectrum features.Furthermore,in order to fully utilize the multispectral information of the images,two multispectral extended forms of the GMABI are constructed simultaneously.The experimental results show that the GMABI and its multispectral extended forms can effectively improve the building detection accuracy.(3)A road extraction approach is proposed by jointly using of multi-features to address the sensitivity of local statistical features of morphological attribute profiles to the semantic environmental changes(rotation,shift,illumination or occlusion,etc.).Includes rotation invariant attribute profile features based on uniform-rotation invariant local binary pattern on attribute profiles,to attenuate statistical differences between same classes caused by variation of direction and illumination.And spatially invariant feature based on multiscale isotropic filtering and spatial aggregation technique is used to enhance region consistency.Finally,a combination of multi-directional Gabor filtering and non-maximal suppression strategy is used to extract road centerlines for the extracted road regions.(4)A road extraction method from HSR images based on the combination of multiscale directional morphological features and graph cut technique is proposed.In order to take full advantage of the salient geometric structure feature of the road networks,a multiscale directional morphology operator is used to HSR image filtering processing to construct spectral-spatial features.To address the shortcoming of the previous methods that are difficult to obtain smooth and coherent road regions,an image fuzzy classification method based on probabilistic support vector machines is introduced.On this basis,a probabilistic propagationbased map-cut road extraction method is proposed.Aiming at the shortcoming that the previous approaches are difficult to obtain smooth and coherent roads,a fuzzy classification based on probabilistic support vector machine is introduced.On this basis,a graph cut road extraction method based on probability propagation is proposed.Finally,a tensor voting method based on perceptual clustering theory is used to adaptively fill the fractured road segments and extract the road centerlines by non-maximal suppression.The proposed method is able to obtain a smooth and coherent road network with no edge burr.