Classiifcation Research Based On Non-imaging Hyperspectral Data

The hyperspectral data’s Multi-band characteristic makes the possible of classification the species.In this thesis, subtropical common tree species in leaves of non-imaging spectral data for speciesclassification of In this paper, we use the hyperspectral non-imaging data of subtropical common treespecies to classification the species, select the data include holly, chestnut, slash pine, cedar and otherspecies, achieving the purpose of improving the signal-to-noise ratio, and band selection. By differentmethod of the tree species classification and band selection, we can achieve the purpose of datadimensionality reduction, and band selection. This has important implications for the future of imaginghyperspectral data’s dimensionality reduction and band selection.In this paper, the absolute mean of the regression residuals for the hyperspectral data. On the basisof the median filter, the first derivative pretreatment, and Select the band by the regression residualsanalysis method, using spectral matching method to classify, we can improve the classification accuracysignificantly. Test accuracy of93%, about13%higher than conventional methods based on the samepretreatment, and this method reduce the data substantially, we can know This method is promising.When the band reaches a certain number, the classification accuracy will decline, this shows that, for aparticular classification, the band number is not as many as possible, some bands not only fail toprovide useful information, but will play a role in noise, these data will reduce the classificationaccuracy. The regression residuals can be removed these unfavorable bands.We investigate the species classification method based on sort the reflectivity of leaf samples,although the accuracy did not improve, but this method throw is worth to continue to explore. But forthe mean compression method for species classification, classification accuracy is not drop by meancompression of the sort of the reflectivity. In the case does not affect the final classification accuracy,the mean compression method can achieve data compress.This article achieves the comparative study of the species classification based on comparison of thenon-processing and value+Differential pretreatment, comparison of the full band and some bands, andcomparison of modeling sample and test sample. At the conventional full-band spectrum matching theclassification, non-pretreatment and median+differential pretreatment classification accuracy are notsignificant differences. In the classification of the regression residuals, non-pretreatment and pretreatedaccuracy is much different, at least a difference of13%. On the basis of the median+differentialpretreatment, the precision of the method of regression residuals was significantly higher thanconventional methods. Pretreatment based on the regression residuals method can improve theclassification accuracy, while compressed data significantly, this method is promising. The analysisshows that the species identification of the classification accuracy will be reduced when the step value isgreater than100. The band selection can reduce the data redundancy, making less data significantly,achieving band selection for hyperspectral data by using residual ideological, The result shows that theabsolute mean method not only can achieve the purpose of band selection but also can improve thespecies classification accuracy.