Extraction Of Regional Residential And Water Area From Remote Sensing Image Based On VWRD

With the development of computer technology, through the high-speed computing capability of the computer, by the rules and logic programmed, using machine intelligence to replace human beings to achieve certain dull and time-consuming work has become possible. Monitoring of land and resources is strong regular, highly repetitive and low-intelligent. Giving such work to the computer can free up a lot of manpower, improve work efficiency, so looking for an automatic update method becomes increasing attention. At present, most update work has been done by operators, accounting for half of the workload as a whole or even above 80%. Enhancing the degree of intelligence and automation of residential area and water updating is the urgent problem to be resolved.Due to the facts that the objects are complex and diverse, the external environment is variable, the same object in different remote sensing images often shows different characteristics. One of the solutions is to divide an object into several sub-objects, each sub-object is simple enough to be modeled. However, the procedures of establishing models and calculating the model parameters are often quite complicated. A new approach to the task of object detection is presented recently, named VWRD(Visual Word Regional Detection). The method is based on the visual word paradigm which is a new introduced concept that has been successfully applied to scenery image classification tasks. In this paper, each visual word corresponds to a sub-model, according to the training samples, building sub-models is convenient and effective. However, as many differences exist between text and images, how to apply VWRD to detect polygon resident area and water, still requires a large number of experiment and in-depth study. The main contents of this dissertation include:VWRD method is introduced and analyzed, and than applied to extract polygon resident area and water from remote sensing images. Key concepts in text recognition are analyzed and the corresponding interpretations in remote sensing objects recognition are given. The usefulness of main theories and technology in text recognition has been checked in the case of resident area and water extraction based on remote sensing image. The learning mechanism of VWRD is discussed, including the learning of visual word bag and the semantic learning of visual words.A new word-building grammar based on visual letters is proposed. Compared with the traditional grammar, it has several advantages, such as the ability of expressing the feature differences between objects in every feature space, high accuracy of extraction, having definite meaning, diversity in the structure of visual word, good extension for visual word. Take the experiential model and its effect in remote sensing image classification for example, how to translate experience knowledge into the meaning of visual letters is introduced. However, two basic problems of the proposed grammar are concluded:how to generate visual alphabet and how to make visual word using visual letters. As far as the two main problems of the visual-letters-grammar are concerned, fuzzy clustering algorithm based on entropy function is applied to divide the feature space in the best way. According to the characteristic of application, the fitness function and mutation procedure in code length are proposed, based on which, the famous Genetic Algorithm(GA) is used to establish the visual word more accurately and effectively.The same visual word usually has different meanings according to its context. So how to use its context information to get the right meaning of the visual word is a hot topic in the field of image segmentation and object recognition. Two methods are applied to translate the context information into a suitable form so that the computer can recognize and handle it. One is visual phase model, which is introduced only in VWRD and used in scene classification; the other is Markov Random Field model, which is the most famous model to express the relationship between the object and the objects nearby. In this paper, both of the two models have been used, based on which, context information has been integrated into VWRD to define the meaning of visual words.Experimental results based on relative algorithm are compared. According to the method proposed in this paper, a change detection system has been developed. During the experiment of the system in different test areas, such as Guangzhou, Wuhan, Beijing and Sichuan, the system has been applied and developed.