Object-oriented Building Extraction From High-resolution Remote Sensing Images Based On Visual Attention Mechanism

With the development of aeronautics and astronautics, the spatial resolution of remote sensing images is increasing, and the difficulty in obtaining data is reduced. Today, with the growing popularity of remote sensing techniques, the automatic extraction of ROI from high-resolution remote sensing images has become research focus. Buildings, which play an important role in human daily life, mark the urban development. Currently, the automatic building extraction from high-resolution remote sensing images has become an important means in urban studies, such as urban sprawl, urban planning, urban heat island effect, population estimation and damage evaluation.Based on visual attention mechanism and object-oriented image analysis method, combining the bottom-up primary feature extraction with top-down empirical knowledge guide, we proposed an object-oriented building extraction method from high-resolution remote sensing images. The main works of the dissertation are listed as follows:(1) Research on color consistency process of remote sensing imagesThe uneven illumination of remote sensing images will make the image interpretation difficult. To eliminate the effects of uneven illumination and increase the accuracy of building extraction, we proposed a wavelet domain enhancement algorithm in HSV color space considering hue consistency. First, the remote sensing image was transformed to HSV color space. Second, the intensity components and saturation components were decomposed by wavelet transform, and an enhancement was implemented. Particularly, to remove the color and luminance differences within optical remote sensing images affected by thin clouds, a customized image dodging algorithm was proposed.(2) Research on objected-oriented multi-scale segmentation of high-resolution remote sensing imagesMulti-scale segmentation plays an important role in objected-oriented image analysis field. In the dissertation, we focused in how to describe the primary features of remote sensing images, and how to use multi-feature in objected-oriented multi-scale segmentation. Base on the time-frequency analysis, we proposed a texture description method of remote sensing images and a measure for textural heterogeneity. And then, we proposed merge criteria based on edge intensity. Combining with heterogeneity criteria, a multi-feature and multi-scale segmentation algorithm of remote sensing images is presented.(3) Research on object-oriented built-up areas extraction based on visual attention mechanismBuilt-up areas can be used as scene information to extract building. First, we analyzed the textural characteristics of built-up areas in high-resolution remote sensing images, and proposed a texture description method of built-up areas. Second, we proposed a Built-up Areas Saliency Index (BASI) based on information theoretic visual attention model. In the calculation of BASI, the texture of built-up areas was used as primary feature. Finally, an object-oriented multi-scale segmentation was performed considering the texture of built-up areas, and built-up areas were extracted according to BASI.(4) Research on object-oriented shadow extraction from high-resolution remote sensing imagesIn high-resolution remote sensing images, building shadow information is an important cue for building extraction. We analyzed the characteristics of shadow in high-resolution remote sensing images, and built a knowledge base of shadow. Then, an object-oriented analysis method was used to segment image in HSV color space. Finally, criteria were established to extract shadow based on knowledge base.(5) Research on object-oriented building extraction based on visual attention mechanismWe proposed a building extraction method combining bottom-up primary feature with top-down empirical knowledge. In high-resolution remote sensing images, buildings usually have high intensity, strong edge and clear texture. To generate the primary features, we proposed a feature space transform method considering building characteristics. And then a Building Saliency Index (BSI) was proposed combining textural feature with spectral feature. The spectral feature in BSI was calculated from feature space transform. To extract buildings, a multi-feature and multi-scale segmentation was performed, and BSI was calculated. Then building objects were extracted by BSI aided by shadow information and built-up areas information. Finally, building objects with smooth edge were obtained by morphological operation.