Tree Species Identification Based On Measured Spectra Of Leaves And Canopies

The identification of tree species is a problem which has a long way to go in forestry remote sensing.It can affect the effects of remote sensing applications especially in forest management and forest ecological assessment field.Hyperspectral data has so much continuous narrowband information that it contain the potential of distinguishing features' subtle spectral differences and make it possible to distinguish tree species with similar spectral features.The emergence of hyperspectral remote sensing technology laid the foundation for solving the problem of tree species identification and it has become a new research hotspot to explore a suitable method for identifying hyperspectral tree species.Since the environmental complexity of different study scales is different,the spectral curves of trees are also affected by environmental disturbance.Therefore,the spectral curves of the same kind of trees are different in leaf and canopy scales.In theory,the accuracy of identifying the same tree group at canopy and leaf scales is different with the same identification method.If a suitable method cannot be found which can make the accuracy of tree species identification on leaf,canopy and remote sensing scale meets the requirements,then it is not feasible to realize the technology route of tree species identification from "leaf to canopy to remote sensing scale".In this paper,four common tree species in the south,such as Chinese fir,masson pine,castanopsis carlesii and geranifolia,were taken as examples to discuss their spectral characteristics and tree species identification methods at leaf and canopy scales respectively,in order to find out whether there is a scheme suitable for tree species identification at different scales at the same time.Three steps are generally required to tree species identification by using hyperspectral data,including,data preprocessing,effective feature extraction and tree species identification.In this paper,four kinds of mathematical transformations,such as first-order differential,second-order differential,logarithmic differential and normalized differential,spectral index and spectral characteristic parameter extraction are used in thepreprocessing step,analyzing the spectral characteristics of each tree species with different pretreatment methods of leaf and canopy scales;the method of principal component analysis(PCA)is used in feature extraction step,exploring the influence of hyperspectral data on the contribution rate of cumulative variance to the rate of increase of principal component number and the relationship between the contribution rate of cumulative variance and the accuracy of tree species identification;BP neural network,SVM and decision tree-CHAID were selected for tree species identification step.Combine the methods of the above methods and get the best recognition solution,exploring which identification schemes have high accuracy and are suitable for both leaf and canopy scales.The main research results of this paper are as follows:(1)The spectral differences between the leaf and canopy scales of the same tree species are mainly determined by the measurement environment and the measurement background.The preprocessing method using logarithmic differential and normalized differential can eliminate the environmental background noise and improve the separability of different tree species.(2)The principal component method can extract the spectral features of the spectral data after mathematical transformation well.However,for spectral data extracted from spectral indices and spectral characteristic parameters,feature selection is not effective.(3)In principal component analysis,the recognition accuracy is generally high when the accumulative variance contribution rate is 90-95%.(4)Normalized differential preprocessing-principal component analysis to extract spectral features-decision tree-CHAID classification algorithm and Logarithmic differential preprocessing-principal component analysis to extract spectral features-SVM classification algorithm are the best ways to tree identification.