Research On Wuxi New District Information Extraction Based On Worldview2 Image

With the rapid development of aerospace technology, the resolution of remote sensing image is becoming higher and higher. The problem of identifying and extracting information efficiently and accurately from remote sensing image is of great importance. It can not only promote the development of high resolution image processing technology, but also has important practical significance for our country’s urbanization and modernization.Object-oriented technology is proposed based on image spectral information, and also takes image objects’ geometric information and texture information into consideration. Because of the feasibility of high resolution image and the ability to improve the accuracy of classification and overcome the "salt and pepper" effect in traditional pixel-based process, Object-oriented technology is becoming a hot field in the study of remote sensing. Research of making full use of all kinds of data, meanwhile removing all kinds of complex interference factors, improving the degree of automation and refinement, has become the main challenges in remote sensing. In this paper, we study the key process in object-oriented classification, which is selection of the optimal segmentation parameters, image objects features extraction and selection of optimal features subset, and the classification effect of different classifier.WorldView2 image data of Wuxi new district is used as experimental data in this paper. The most widely used segmentation method, fractal network evolution algorithm, is used in this paper. To solve the problem of optimal segmentation parameters selection, this paper proposed a model, which take objects’homogeneity and heterogeneity into account, and used the model above to evaluate the effects of segmentation under different parameters using objective, quantitative Criterion. This article firstly used original band correlation analysis and covariance statistics to determine the weight of each band, then used the model to evaluate segmentation effects of 63 groups of different segmentation scale factor and spectral parameter combinations, and finally find the most suitable parameters using global optimal segmentation quality evaluation model. The segmentation parameters selected by this article using the model above is considered of good effect through artificial visual inspection.For the problem of "dimension disaster" in object-oriented classification process, study of eliminating invalid features is of importance. In this paper, the class of building is divided into four subclasses based on spectral analysis. Various features of image objects, like Spectral features, geometric features and texture features, have been extracted to fully extract high resolution remote sensing image’s features information. This article then use random forests algorithm to evaluate 83 features importance. Moreover, this paper use recursive feature elimination strategy to determine the candidate optimal features subset dimensions. On the basis of the process above, this paper use the correlation analysis to remove redundant spectral features to get optimal feature subset consist of 12 characteristics. Through comparing image objects’separability on part of optimal features subset and original spectrum features, the result shows that method used in this article is very effective, and the separability of different classes has been greatly improved by using feature extraction and selection described above, thus verify the validity of method used in this paper.In this paper, on the basis of selection of optimal segmentation parameters and feature extraction and selection, we used 3 different classification methods to study the object- oriented classification process of high resolution remote sensing images, which are KNN, SVM and RF. Moreover, confusion matrix were used to analyze and compare the results of 3 different method mentioned above. The results show that object-oriented method can extract different classes’ information reasonably meanwhile effectively avoid the "salt and pepper noise" phenomenon which is common in pixel-based classification. By comparison of accuracy of 3 object-oriented classification used in this paper, we can conclude that Random forests are better than other 2 methods. RF method has higher classification accuracy due to its effective separation ability of building information and other impervious. At the same time, due to RF’s advantages of high calculation efficiency, simple parameters setting, it will be widely used in the field of object-oriented classification in the future.