Modified Nearest Feature Space Approach to Hyperspectral Image Classification

• Main Authors: Tzu-Wei Tseng, 曾咨維
• Other Authors: Jyh-Perng Fang
• Format: Others
• Online Access: http://ndltd.ncl.edu.tw/handle/hs3e3f

碩士 === 國立臺北科技大學 === 電機工程研究所 === 104 === In recent years, the rise of "big data". Data analysis and processing are gradually popular, and data classification is the most important among them, and then the most of classification algorithms were developed. For example, Nearest Feature Space (NFS) was used in classification of hyperspectral remote sensing data. The main idea of NFS algorithm is using the shortest distance from test point to Feature Surface (FS), as the basis for classification. Three training point of each FS have information to improve overall accuracy. However, when the distribution of sample points and test point are too close, NFS algorithm would be difficult to categorize and easily cause misjudgments. To overcome aforementioned problem, this paper proposes the new method which is called “Modified Nearest Feature Space” (MNFS). Through analyzing the coverage of the FS, to limit the extensible range of each FS by the analysis result so as to reduce the impact of the overlapping between each category, and then it got better classification accuracy. Finally, the experiment prove that MNFS is better than NFS.