Combination And Comparison Of Dimensional Reduction And Classification Algorithms For Classifying Main Tree Species Of Southern China Base On Hyperspectral Data

The development of hyperspectral techniques has made it possible the identification and classification of tree species, while they are often very difficult using traditional multi-spectral imagery of remote sensing. But, because of high dimension, how to achieve recognition and classification of tree species using hyperspectral data with high efficiency and accuracy and to further develop new theories of remote sensing application and classification algorithms are a great challenge. Based on the analysis of existing hyperspectral classification algorithms, this dissertation aimed at deeply and systematically studying and developing dimensional reduction and classification algorithms of hyperspectral data for tree species of southern China.In order to seek the hyperspectral bands that are sensitive to distinguish main tree species of Southern China and to further develop appropriate algorithms of dimensional reduction and classification algorithms, reflectance observations of four main tree species including Cunninghamia Lanceolata, Pinus massoniana Lamb, Cinnamomum camphor a and Liriodendron chinensis were collected at fixed points and times for more than six years,and more than3000spectral data sets were collected. After processing, a total of310spectral data sets were obtained. At the same time, the levels of chlorophyll, carotenoids and xanthophyll of branches and leaves for the corresponding trees were measured.Using the aforementioned data, four dimensional reduction methods including Principle Component Analysis (PCA), Independent Component Analysis (ICA), combination of Genetic Algorithm and Support vector machine (GA-SVM), and combination of Particle Swarm Optimization and Support Vector Machine (PSO-SVM), were analyzed and compared. Moreover, correlation analysis of the ground spectral reflectance data and the observations of main pigment concentrations of the tree species were conducted. The dimensional reduction of the spectral data collected in the field was carried out. Using the data after dimensional reduction, five classification algorithms, including Support Vector Machine (SVM), back propagation (BP) neural network, Mahalanobis Distance classification, Bayes classification, and Spectral Angle Mapper (SAM), were compared based on classification accuracy. This led to the best combination of hyperspectral data dimensional reduction and classification algorithms for these tree species using the ground spectral observations. Finally, these combinations of algorithms were validated using Hyperion imaging spectrum data. The main findings were as follows:(1)It was also found out that two data transformations including logarithms transformation and logarithm first derivative transformation out-performed others for dimensional reduction and classification.(2)Based on comparison of dimensional reduction algorithms and selection of bands using ground hyperspectral observations, it was found out that PCA was better than other algorithms for dimensional reduction and classification of tree species.(3)It was found out that the best band combination for classification of tree species was obtained from403nm,457nm,505nm,517nm,529nm,538nm,553nm,622nm,667nm,679nm,703nm,730nm,763nm,853nm.(4)Based on the correlation analysis of the ground spectral reflectance observations with the main pigment concentrations of the tree species, it was found out that the best band ranges were401-504nm and659-686nm, which resulted in the highest classification accuracy of more than90%. There are smaller correlations between the ground spectral reflectance observations with carotenoids and xanthophyll concentrations.(5) The results base on Hyperion imaging data that have a spatial resolution of30m X30m indicated that the bands selected using GA-SVM were more appropriate for the classification of the tree species for Southern China.