Typical Tree Specks Classification Study Based On Leaf Hyperspectral Data In Jiaohe,Jilin Province

Leaf spectral features closely related to tree species information are the foundation of the tree species classification using hyperspectral remote sensing data. In this paper, the leaves of the main tree species in the experimental forest are studied to explore the probability of extracting the competition situation of the tree species in mingled forest using satellite data, based on the laboratory measured leaf spectral data of broad-leaved Korean pine forest in Jiaohe, Jilin Provience. The experimental data is obtained from three stages. In the first stage, the reflection and transmission data are gathered respectively from both ventral leaf surface and dorsal leaf surface of the nine dominant tree species. The data is pre-processed by three methods:logarithmic, first-order derivative and second derivative. In the second and third stage, the data is obtained from twelve and eleven main tree species, which is pre-processed only by the best method first-order derivative. Then the data is used to analyze the classification accuracy utilizing four methods, including ANOVA, discriminant analysis, factor analysis and clustering analysis. And we studied how hyperspectral of leaf change under different space or time.The results indicate that the classification of needle-leaved and broad-leaved tree species is perfect with an accuracy of 100% using the leaf spectral data obtained from the experiment. The classification accuracy of nine tree species is high to 80%～100%. The classification accuracy of twelve tree species is generally high with 63.5% of the overall classification accuracy located between 90% and 100%. First order derivative data consisting of visible and near-infrared is the best for classification.19 feature bands were filtered and selected for classification, which are 494,502,642,650,654,662,682,686, 682,682,730,734,838,734,734,1260,1532,1536 and 1532 nm respectively. Classification accuracy of feature bands can reach around 70%. In consideration of the low spectral resolution of high spatial resolution satellite, the spectrum of spectrum sensor GEOEYE-1, RAPIDEYE and WORDIVEW2 is simulated based on the first stage data and band response function. Then the simulated spectrum is used for the study of tree species classification. The results show that simulated spectrum can classify the needle-leaved and broad-leaved tree species effectively and the classification accuracy is in 71.6%-100.0%, but for 9 tree species the classification accuracy is very low which is in 47.3%-74.0%. The results also show that the spectra of different tree species vary with elevation. Among them, Acer mono, maple and aspen have shown the biggest change, while tilia amurensis the smallest. During discoloration period, amur cork tree is the specie that can be easily separated, while acer mono, maple and tilia amurensis are the most difficult to be classified.Experimental results show that hyperspectral data of species in Changbai mountain area has high separability. It reveals that the band number of sensors used should be more than eight when classifying tree species using hyperspectral data. At the same time, band range should cover the visible to near-infrared (NIR). First derivative of the data will be better for classification. Habit of tree species should be considered because leaf hyperspectral changes with time and space.