Application Of Object-oriented Classification Based On SPOT Image

With the development of remote sensing technology, remote sensing image has a higherspatial resolution and makes us obtain very abundant information of nature.In addition to thespectral information outside of itself, it also contains a large number of shapes and textureinformation. The traditional image classification method that based on spectral of pixels, canworks well for the image that have middle or low resolution. But this method cannot be fullyutilized other information such as texture, not well adapt to the development of today’s remotesensing data and classification accuracy. In this context, the classification of object-orientedmethods have emerged. This article will focus on method research and application ofobject-oriented classification.In this paper, the objected-oriented classification method has been studied deeply. Andusing the SPOT5in Baiquan County, Heilongjiang province as data sources, usingpixel-based classification and object-oriented method for classification, comparison of twokinds of qualitative and quantitative classification results separately. The main contains andresults are as follow:(1) Comprehensive use ISODATA and maximum likelihood classification for thetraditional classification method. First, obtained an initial classification of spectral similarityof spectra of cluster group using unsupervised classification. With this result, select a certainnumber of training samples for each type of object in the image to maximum likelihoodmethod. When sampling, combined with the interpretation tables for accurate selected. Postclassification processing after the initial results of the optimization results.(2)In object-oriented classification, first for image segmentation. Different features hasthe different best scale. Using visual methods to determine best scale for each layer andfeatures suitable for extracting, through a series of segmentation test.(3) Fuzzy classification contains two major classifications: member function and nearestneighbor classification. Membership function is suitable only for which can be classified withother features using one or several characteristics. Nearest neighbor classification applied forthe features that cannot be easily distinguishable with other that use only a small number ofparameters. Using quantitative methods to select optimal feature space for the nearestneighbor classification, and in this article includes the NDVI, mean layer, shape index of theseven characteristics.(4)Compare the two kinds of classification results both in qualitative and quantitativeaccuracy. From the visual aspect, images obtained using traditional pixel method is broken,there is a "salt and pepper", while object-oriented classification results obtained betweencontinuous and strong, more realistic. Quantitative evaluation of precision, pixel-basedmethod for overall accuracy is0.6667,Kappa coefficient is0.6123;classification accuracy ofobject-oriented result is88.17%, increased21.5%than the overall accuracy of the method based on pixel (66.67%), Kappa coefficient is0.8605, increased0.2482than the methodbased on pixel (0.6123). Indicates that object-oriented classification method is effective inimproving results, improving the accuracy of classification.