Optimal Bands Selection For Tree Classification Based On Hyperspectral Data

Hyperspectral is known as its high spectral resolution, hyperspectral remote sensing as a hot issuehas become an international remote sensing technology. The rich spectrum information supported byhyperspectral remote sensing which provides convenience for the classification and recognition of theland covers. Currently the technology of hyperspectral remote sensing has been utilized in many fields.While from now on there is not a much better classification method wich used for the treescalssification. Several different band selection methods were analyzed in this paper, and theclassification accuracy of the different methods was compared. The study is focus on the bandsselection methods, in order to explore ways to improve the classification accuracy.Considering choosing the bands whose spectral reflectance varies smaller between the differentsamples for one class as the optimal bands, bands selection method based on regression residual wasused in the paper. In addition, considering the theory of Sequential Backward Selection, another methodof Sequential Backward Selection based on regression residuals was proposed. These methods and theSequential Backward Selection method based on deviation between groups within groups were used inthe tree species classification based on leaves non-imaging spectral data and EO-1Hyperionhyperspectral data. Both the results of non-imaging spectral data and EO-1Hyperion hyperspectral datashow that the method of Sequential Backward Selection based on regression residuals has a betterclassification than that of the bands selection based on regression residuals. Meanwhile the FeatureExtraction method of Partial Least Squares Regression was also studied in this paper, and proved be abetter method.In the study of the bands selection for the EO-1Hyperion remote sensing image, five methodswere evaluated, including Sequential Backward Selection, Stepwise discriminant analysis, geneticalgorithm and TM-bands method. The results show that the classification accuracy of testing sampleswith bands subset is superior to that of the whole bands. We found that stepwise discriminant analysisand the genetic algorithm method are more effective in our experiments, with a classification accuracyof74.73%and73.69%respectively, and TM-bands method is followed whose classification accuracy is70.06%. In contrast, sequential backward selection method has relatively poor performance, for itsclassification accuracy lower than TM-bands method, while whose bands number is five times largerthan TM-bands method. In addition, Fisher Discriminant Analysis was studied for EO-1Hyperionhyperspectral spectrums reduce. And the classification accuracy for the testing sample is78.24%whichis higher than the other bands selection methods mentioned above.The classification results show that with a few bands the classification accuracy can reach or evenhigher than that with the whole bands. Therefore, using the appropriate bands selection method we canget a even better classification accuracy in the study of tree classification with hyperspectral. At thesame time we can improve the efficiency of data processing for the data is much compressed.