The Study Of Spatial Scale Effects In Typical Objects' Feature Extracted From Remote Sensing Image

With the development of remote sensing technology, the applications of remote sensing are constantly expanding. However, the existing data can not meet the need of all studies. The smaller features can't be clearly expressed in low spatial resolution images.And the relationship between the positional and the feature of the ground object is fuzzy. The feature category of the high spatial resolution images has some characteristics, including a significant structure, a clear location distribution, obvious texture features and size of the shape, etc. But the high spatial resolution images have large volumes of data processing and information redundancy, which results in a waste of resources. The complex geomorphological features and geographical structure can't be accurately expressed by a single scale of remote sensing image, And the detailed information of typical objects described by different scales of image from different angles are also different. In a certain scale remote sensing image, the information of a certain type of feature category can be expressed accurately, but it is not necessarily to the information expression of other feature categories. This is because that the complexity of the shape, texture and internal information of various objects in the scale images are also different. In view of the extraction of the typical features, when the spatial resolution is different, and so it is with its object features.But in reality, it is not possible to obtain the continuous spatial resolution images directly. Therefore, a scale conversion to the existing high spatial resolution images is of great significance. The work such as making reasonable use of the relationship between the scale of remote sensing images and object feature extraction, comparing and analyzing the study objects of the same area to calculate the most suitable image expression scale helps to reduce the waste of resources and eliminate the problem at the maximum degree of data redundancy caused by low efficiency.Taking Ludian County as the research area, this paper selects World View-2 images with the resolution of 0.5m as the scale conversion data source, and Landsat-8 and GF-1 images as the classification and comparison data. Calculating a series of images which has been conversed and analyzing the appropriate expression scales of the image in the study area; Selecting five typical features sample, including paddy field, dry land, forest land, construction land and water, and then calculating the sample separability, so the most appropriate scale expression typical objects will be obtained. The specific research contents and results are as follows:(1)A series of scale images which have been converted by cubic convolution interpolation method that are evaluated with four scale effect indexes: Pixel mean, standard deviation, information entropy and variance. The result indicated when the spatial resolution is 12m-24 m, the remote sensing image keeps better spectral characteristics of the initial image, it also has good clarity and the information it contains are more abundant. The remote sensing image like that can be used as the appropriate scale data of study area.(2)Five typical samples of different scale images in the study area are selected, and the changing process of shape, profile, spectrum and texture features of objects with spatial resolution are visually displayed. At first, to calculate the sample separability, and then,to statistics out the best separation scale of each typical features, and finally the conclusion was drawn out: the optimum water extracting spatial resolution is 30m; the best resolution to distinguish paddy field and other four kinds of objects is 12m; the best spatial resolution to express the dry land and forest land is 16m; the construction land which was at a spatial resolution of 15 m can be better separated.(3)Finally, the SVM(Support Vector Machine) classification method is used to classify the data of Landsat-8? GF-1 images and the series of scale converted images, and then to evaluate the overall accuracy and kappa coefficient of all the classification results. It is concluded that the classification accuracy of the resolution of 15 m and 16 m images is significantly higher than that of the resolution of 15 m of Landsat-8 and GF-1 images; According to the accuracy evaluation and analysis of the series of the converted image classification results,that the image classification results are better, and the best resolution is 16 m, which can be used as the resolution of the typical features in the study area.