| Summary: | 碩士 === 國立嘉義大學 === 森林暨自然資源學系研究所 === 106 === Classification of tree species is necessary at the level of managing forest resources. How to use forest images acquired by remote sensing technology to do accurately classify for tree species in forests is a very important challenge. In Modern remote sensing technology, multispectral and hyperspectral sensors can provide more spectral information and detect a variety of surface information. In this study, take semi-developed state of forest area in the middle and low altitude for example. Using the CASI hyperspectral image and the tree species information samples base on the ground survey to do the supervised classification, and then to investigate the effectiveness of different classifiers and image preprocessing methods for mapping land cover types and tree species maps for CASI hyperspectral imagery. The research results show that the pre-band screening can effectively reduce the data dimension and facilitate the segmentation of objects, to construct the characteristics of spectral homogeneity in spatial clustering, which is conducive to the classification of land types and tree species. Using the support vector machine classification (MNF-SVM) of MNF pre-processing image object segmentation, the best classification performance of terrestrial type can be achieved. The overall accuracy and Kappa agreement coefficient (Kappa) are 95.33% and 0.95 respectively. However, both MNF-SAM and MNF-SID can achieve OA=93% and Kappa=0.92 or higher. The three kinds of image pre-processing methods have the best MNF, the ICA is the second, and the PCA is the smallest. The MNF-SVM is better than the PCA-SVM in the classification of tree species. The Kappa of each is 0.95 and 0.81.
|
|---|
