Study On Application Of Object-oriented Analysis To Earthquake Damage Extraction From Remote Sensing Image

After earthquake occurred, accurate and rapid evaluation of the disaster damage is urgently needed for decision-making of emergency commands and rescues. Remote sensing information is superior to conventional ground survey in such aspects as faster, more stable, less ground condition limits. Recent years, remote sensing earthquake damage evaluation has been one of the major methods, and naturally auto-detection of earthquake damage of remote sensing image has become the critical step in the action. The classification precision directly influenced the whole work result. Therefore, it is important to keep the stability and to improve the precision of the auto earthquake damage detection.The high spatial resolution remote sensing data has much more complicated texture and less spectral information than other data. With the fast development of high spatial resolution remote sensing techniques, the traditional pixel-based classification method has been far from requirements. It demanded us to find other ways to excavate more useful information from the data. Object-oriented image analysis method solved these problems by changing the basic unit from single pixels to image objects. The special multi-resolution segmentation could extract image object primitives in arbitrary resolution. Additional features such as texture, shape, relationships of the image objects were synthetically using with the spectral information. The method simulated visual interpretation, and adopted fuzzy-logic classification method to accomplish the effective detection.Object-oriented image analysis method has been successfully applied in various fields. But the application in earthquake damage extraction is still immature. In this paper, further study in earthquake damage extraction of remote sensing data has been realized.Firstly, this paper analysed a great deal different sourses data of the Ms8.0 Wenchuan earthquake which occurred at May 12th, 2008. The research proceeded a series of qualitative change detection and visual interpretation. The results revealed kinds of damages such as geological hazards and building damage in multi-scales. And for precision evaluation, we also conducted a field study to obtain abundant detail data.Secondly, on the basis of traditional earthquake damage analysis, object-oriented image analysis method was introduced into the research. A further exploration of the principles was dissected, and study on selection of the parameters and features was carried out.There are two major parts of the object-oriented image analysis method: image segmentation and fuzzy-logic classification.Image segmentation constructs a hierarchical network of image objects. The process is able to find image objects in any chosen resolution between the image and pixel levels. Spatial resolution influences the scale choosing, and suitable scale parameters (image layer weights and composition of homogeneity criterion) satisfy particular requires of spectral or texture information.Fuzzy-logic classification is based on image segmentation, and the image objects are classified as original units. Instead of dividing things by 0 or 1, fuzzy mathematics divides things by the interval [0, 1]. A second semantic network, the class hierarchy, is built up in this part, and the information transmission between classes comes true. The advantage of fuzzy-logic is the possibility to integrate most different kinds of features and to connect them by means of fuzzy- logic operators.Thirdly, three areas - Beichuan, Longmenshan, and Dujiangyan - were chosen to practise the research for different remote sensing data and various aspects of earthquake damage.Beichuan county and its surrounding area was used to fast extract earthquake damage area. Through application of two segmentation levels and simple spectral features, the damage area was extracted from the remote sensing data, the spatial resolution of which was 8 m.Longmenshan town in Pengzhou was chosen aiming at rural building damage extraction. Low buildings are the most common building type in rural China and the buildings always spread dispersedly. Therefore the research chose some simpler classes and fewer scales. The object-oriented classification of Longmenshan was carried out by using of single scale level and spectral features.Dujiangyan was chosen for city buildings damage extraction. Considering the complicated and uneven characteristics of the city buildings, the research first chose a small cut of the image as sample. The resolution of the remote sensing data is 0.5m, and four segmentation levels were built upon the data to classify. During the process, kinds of features were combined together, and it contained spectral, shape, and relationship features. Traditional supervised classification based on pixels was also operated as control.Finally, integrated all researches above, some conclusions were gained as follows:1. Object-oriented classification of remote sensing data is suitable for fast earthquake damage areas extraction. The results could provide evidence for emergency command and rescue decision making.2. Object-oriented image analysis method could be widely applied in multi-resolution remote sensing data and different areas, such as rural areas and cities, and there is no influence on the classification precision of different areas.3. Compared to pixel-based classification, the object-oriented classification of earthquake damage could eliminate nosies as"pepper effect", and it gets stable results towards different areas. The most important is that, to a great extent, object-oriented method improves the precision.