Research Of Method And Application For Vegetation Classification On Hyperion Image

As a kind of renewable natural resources on land, forest resources greatly satisfythe material resources required of the human survival and development. Vegetation isthe main part of forest resources, accurately identify the vegetation is the basis of researching and utiliz forest resources, which can also provide scientific and reasonable suggestions for forest resources management.The developed remote sensing technology makes the achievement of the forestvegetation information large-scale, real-time and dynamic monitoriing. However, thecommon multi-band remote sensing can not detect the tiny difference of differentvegetation spectrum. Hyperspectral remote sensing has broken through this limitationit can get hundreds of continuous narrow-band spectral curve from visible and nearinfrared to shortwave infrared band, which can reflect the the unique spectral featureof different vegetation and get the diagnostic spectral. So hyperspectral remotesensing has large advantages at classification of forest vegetation. This paper takehyperspectral image as the data resource to carry out the classification research,several results gained below:1. The hyperion image data that this paper chooses has many ‘non-normalpixels’ or poor-quality bands. To guarantee the accuracy of information extraction,aseries of preproccess must be made, which includes radiation values conversion, bandexcluding, bad lines repairing, removal of vertical stripes and Smile effect,atmospheric correction. Radiometric calibration can get the real reflectance spectralvalues; Band excluding has removed the uncalibrated bands and the bands affected bywater vapor; Bad lines repair dealed with the lines of which the values are0or near0.Stripes removal repaired the bands that has noises. Smile effect removal got rid of theSmile effect of VNIR band. The land objects’ real reflectivity was got afteratmospheric correction. After preproccess, the Hyperion data’s data quantity reduced,the images’ quality increaced and the the vegetation spectral curves tend to be realspectral feature.2. High-dimensional spectral data of hyperion image data can easliy lead toHughes phenomenon during the classification. To improve the classification accuracy, dimension reducing was made to the hyperspectral image data. In addition to thecharacteristics of hyperspectral data, segmented principal component and band indexcombines method were used, which not only solved the correlation problem existedwhen segmented principal component was used to choose band, but also reduced thecomputing work. The final original band combination choose was (33,91,151), on theimage of which the vegetation information feature was obvious which was helpful toincreace the accuracy of vegetation classification.3. The choice of classifier is the key technology in the classification. Based onthe key technology of the object-oriented classification and the basic principle ofdecision tree classification, of which the combined CART algorithm could be used tothe hyperspectral forest vegetation classification research. Firstly, the multiscalesegmentation method of object-oriented was used to carry split test several times onthe Hyperion image. And the segmentation threshold was35. The homogeneousobject gained after segmentation, of which the characteristics were relativelyhomogeneous, was used as pending unit. The feature variable of segmented objectincluding spectral, texture, shape and terrain features was extracted. Covariance andcorrelation were used to screen variables. Finally, the CART algorithm of decisiontree was used to choose optimal variable from the feature variables left after screeningto build decision tree. In the process of build decision tree, the automatical obtainmentof classification rules could be realized and the classification threshold could bedetermined, which reduced the influence of human factors. To objectively reflect thesuperiority of the object-oriented decision tree classification method, the traditionalmaximum likelihood method and neural network were used to classificationcomparison experiment. The results showed that object-oriented decision treeclassification achieved the highest classification accuracy on the forest vegetation.The combination of Object-oriented classification and decision tree algorithmprovided a new idea for vegetation classification of remote sensing images, whichwidely used in land dynamic monitoring and land resources survey...