Study On Object-oriented Remote Sensing Image Classification And Its Application

The development of the remote sensing technology, especially with the appearance of high resolution remote sensing, makes us observe the nature more extently and deeply. By far, the remote sensing technology is applied by extracting the interesting information from the satellite images. But the high resolution remote sensing image such as IKONOS and QuickBird has richer spatial structure information which contains shape and texture information. Although the remote sensing image classification technology has developed considerably, it will result in not only reducing the accuracy of classification but also making the spatial data redundant and wasting the resource when the single traditional classification method based on spectrum of pixels is applied to the high resolution remote sensing image.At present, local governments attach much importance to the urban vegetation construction, and apply the remote sensing technology to the urban vegetation investigation. Although extracting vegetation using low or middle resolution images is rough, it has promoted the remote sensing technology to be applied in the urban vegetation investigation. High resolution remote sensing image has higher spatial resolution and richer information, can display the urban vegetation in a deeper level. So, the high resolution remote sensing now is widely applied to the urban vegetation investigation instead of the low or middle resolution remote sensing.In this paper, the Object-oriented classification method has been studied deeply. Then the detailed vegetation information is extrated from the QuickBird image through the Object-oriented classification method as an example in Futian, Shenzhen City. The article is written for the purpose of understangding the Object-oriented classification method more deeply in theory and application. The main contains and results are as follow:(1) Remote sensing image must be scaled before the Object-oriented analysis. In the paper, the Object-oriented scaling is researched based on the pixed scaling. It segments the high spatial resolution remote sensing image by the multi-scale segmentation technology to produce many levels composed of objects, builds the hierarchical structure of the remote sensing information in different scale, which can make the remote sensing information be transmitted through different levels. And the result of scaling is also evaluated by using Mean, StdDev, Mean of NDVI and fraction dimension for indices. The result indicates that the appraisal indicator values of classes vary with different scales, and the variety tendency of different classes is not same. (2) The scale effect .of remote sensing information extraction is proved to existe through experiments in the paper. It proves that the extraction accuracy is considerably different in different scale, and there exists a theorical optimal scale in which the extraction accuracy will be highest to single class extraction.(3) The optimal scale is the key of multi-segmentation and information extraction accuracy increasing. So, mean area ratio method as a new method for selecting optimal segmentation scale is developed base on the principle of best segmentation results. For this method, to the assumption that the scale with which the best segmented objects are acquired by segmenting image is the optimal scale, a mean area ration index which is for measuring matching between segmented objects and surface objects is defined. There are two factors: one is the ratio of a surface object area and the segmented objects total area which display the matching between the border of segmented objects and the border of surface objects, and the more value of which closes to 1, the more they matches each other.; the other is the number ot segmented objects which represents the dgree of fragmentation, the fewer the segmented objects the lower fragmentation, segmented objects displayes the spatial and shape information of surface objects. This method is proved to be good at selecting optimal scale.(4) The object feature space which determines the distance parameter d and the function slope on the impact of the classifier quantity are studied based on the expression of nearest neighbor model. The feature space with best class distance is determined to be theorial feature space by building class distance matrix. The optimal feature dimension for classification is not more the 10 through the experiment results. The membership curves with different function slope are also different, and the membership with distance flats with the function slope increasing. When the function slope becomes smaller, it stretches membership in the range of small distance, enhances the discrimination of classes with similar features. On the contrary, when the function slope becomes larger, it contracts membership in the range of larger distance, enhances the discrimination of classes with different features.(5) In order to solve the problem of membership function construction, the fuzzy statistical method is developed to construct the optimal membership function close to the real one. To this method, class feature range is discreted firstly. Then the frequencies of selected class feature values are in the range of each interval. At last, a continuous membership function is acquired by fitting these scaterred points composed of the middle interval value and the membership. It is proved that the membership function constructed by this method can improve the extraction accuracy. (6) The fuzzy multi-classifier model is developed through combining nearest neighbor classifier and the classifier based on fuzzy rules by cascading. With this classifier, the classes which are easily to be extracted are firstly extracted by fuzzy rules classifier; the rest mixed classes which are difficultly to be extracted are then extracted by nearest neighbor classifier; all extracted classes are combined. It is proved that the fuzzy multi-classifier model can effectively increase the extraction accuracy.(7) All that the remote sensing image is segmented by multi-segmentation technology through selecting scale parameter using the mean area ratio method and adjusting other parameters manully, class hierarchy is constructed, fuzzy rules and feature space are defined, are done to extracting the vegetation in Futian after image preproceeding such as fusion, geometric correction and filtering. First, vegetation is preliminarily classified as grass, forest, shrubbery and sparse forest, other vegetation. Sencond, forest is detailtedly classified as evergreen broad-leaved, fruit forest, mountaintop elfin forest, mangrove, other forest. Third, the classification accuracy is evaluated, and compared to the maximum likelihood and minimum distance method. It is showed that the Object-oriented classification method improves the extraction accuracy considerably whether in land cover classification or in vegetation information extraction.