Land Use Classification Based On Hyperion Hyperspectral Images

Hyperspectral remote sensing are not only able to provide image data of Space targetcharacteristics in nanoscale, but show dozens or even hundreds of narrow continuous spectral bandsfor each image pixel in a wide wavelength range. Then various land cover types with subtle spectraldifferences in an area can be accurately detected and separated. But, It cannot be denied that there areseries of problems, such as，mixed pixel, high correlation between bands, lower spatial resolution. Sotwo classification methods, pixel-based and sub-pixel-based were used to show application potentialof Hyperspectral data, then provide effective help for Hyperspectral image classification research; andthe more important is to decrease the accompanying influence of problems above to give full play tothe advantages of spectrum to achieve the goal of improving land use classification accuracy.Taking Baizhang town in Yuhang district of Hangzhou city in ZheJiang province as test site,study was carried out on the land features identification using spaceborne Hyperion hyperspectralimages. Three processing steps were included for land features identification. In step1, the studiedsituation of two non-parametric machine learning classifiers based on pixel, include SVM (supportvector machine) and NN (neural network) were compared, with the same Training sample conditions.The results indicate that SVM was superior to NN. In step2, abundant spectral information afterdecomposition of mixed pixels combined with texture information under texture analysis were tried tocarried out to realized the result of classification. The new method mainly meeted the classificationneed. the main conclusions were as follows:(1) SVM (support vector machine) and NN (neural network) have similar net modality andclassification essence, but due to the different requirements of input data samples and differentfunctional structure, SVM obtain globally optimal solution, yet NN acquire only local solution. Theproduction accuracy range and user accuracy range of objects involved in the classification usingSVM was63.89%-98.67%,67.26%-95.39%respectively, while the one of NN was2.95%-99.21%，33.23%-94.18%. And the classification accuracy of Pinus massoniana Lamb and oak using NN is16%and15%higher than SVM. (2) Relied on other assistant properties, texture analysis can play an important role in identifyingland futures of low spatial resolution remote sensing data, especially in distinguishing some finevegetation types. The classification result of spectrum with texture information has improved a lotcompared with the other two methods with single spectral information: the classification accuracy ofbuilding is34.13%and17.16%higher than the one of above two methods, the one of farming is19.71%and9.24%, Pinus massoniana Lamb is27.09%and5.42%, oak is nearly3%and10%. Thecombination of a variety of classifier or the comprehensive use of characteristics was an Effective wayto improve classification accuracy of the fine vegetation types. And the result clearly demonstrateddifferent sizes of texture window generate different influences. The overall accuracy and KAPPAcoefficient show their largest number under the window of5×5, reaching87.32%，0.83.For Low spatial resolution of Hyperion hyperspectral data, mixed pixels are ubiquitous, twonon-parametric machine learning classifiers based on pixel, include SVM (support vector machine)and NN (neural network) can identify the fine vegetation and obtain the satisfied classificationaccuracy, which proved the Spectrum advantage of Hyperion hyperspectral data.