Classification Of Forest Types Using Airborne PHI Hyperspectral Data

Hyperspectral remote sensing has advantage of high spectral resolution,large volume of information,the continuous spectrum of different features and identification of subtle features as a passive optical remote sensing technology,which has been used widely in forestry,ocean,atmosphere,geology and other fields.PHI hyperspectral data was the data source for dominant tree species identification in the paper,considering that the information of hyperspectral data non-sensitive bands for vegetation is low,especially the noise of both ends of the spectrum is significant,adaptive band selection and differential method was used to reduc dimension and noise in the paper.Hierarchical classification strategy was used in order to eliminate the interference of non-forest features for classification,at first according to NDVI threshold to distinguish between forest and non-forest,then the SVM and SAM method was used to distinguish the dominant tree species further.The main results and conclusions are as follows:(1)It removed noise effectively,reduced redundant information and computation by ICA.The quality of selected bands was better than before by comparing the images and spectral curves.(2)NDVI has a good effect to distinguish between forest land and non-forest land.Due to the time of data acquisition was in winter,deciduous trees have leaves and vegetation spectral features were not obvious,the NDVI of deciduous trees was close to shrub grass and the spectral information of tress was close to non-forest(such as farmland or roads),it's very important to distinguish forest and non-forest land before tree species.Study showed that it was effective to distinguish between coniferous and broad-leaved forest from evergreen and deciduous forest,as well as non-forest by hierarchical classification strategy.(3)The accuracy of SVM and SAM was 82.2% and 78.1% by ABS method;The accuracy of SVM and SAM was 75.8% and 73.3% by first-order differential method;The accuracy of SVM and SAM was 79.1% and 77.3% by second-order differential method.It showed that the accuracy of SVM is higher than SAM by comparising two types of classification results with dimensionality reduction methods.In addition,ABS method is most suitable for dimensionality reduction with PHI data,while the second-order differential method was better than the first-order differential for dimension reduction,the accuracy of ABS and SVM was more than 80% in general.(4)PHI hyperspectral data applies to classification where the type of forest is single and forest aspect are kept in good order.It's feasible to distinguish forest land where the type of forest is single and forest aspect are kept in good order,as well as the accuracy of five kinds of results were more than 75%.