A 2-Dimension Impurity Function Band Prioritization Based on Mutual Information for Hyperspectral Images

碩士 === 國立臺北科技大學 === 電機工程研究所 === 103 === In recent years, scholars have proposed to use Impurity Function Band Prioritization (IFBP) to select high correlated bands according to calculating category coverage in one-dimensional. Using this method calculate high overlapping in one-dimensional but not i...

Full description

Bibliographic Details
Main Authors: Shih-Chieh Chien, 簡士傑
Other Authors: Jyh-Perng Fang
Language:zh-TW
Online Access:http://ndltd.ncl.edu.tw/handle/98499c
Description
Summary:碩士 === 國立臺北科技大學 === 電機工程研究所 === 103 === In recent years, scholars have proposed to use Impurity Function Band Prioritization (IFBP) to select high correlated bands according to calculating category coverage in one-dimensional. Using this method calculate high overlapping in one-dimensional but not in two-dimensional. Because sometimes bands selection is lack of information in one-dimensional and is short of relation between bands, Lead to less dimensionas reduction in select priod bands. According to what we mention above in these report we have use these two method which are Particle Swarm Optimization (PSO) and 2D-IFBP. Unlike IFBP is lack of information between bands, increase information to select the most representative band to get dimensionality reduction result in 2D-IFBP.Combine PSO with Greedy Modular Eigenspace method (GME) to cluster the highly correlated bands together and use IFBP method to select representative bands. Next we use Mutual Information (MI) concept of calculating impurity function class overlapping with bivariate normal distribution to calculate correlated bands, then use 2D-IFBP to adjust IFBP’s bands weighting scores statistics result for obtaining greatly dimensionality reduction effect. The effectiveness of the proposed method is evaluated by HYDICE and AVIRIS remote sensing images for testing. The experimental results show 2D-IFBP method could effective select most representative bands and get high classification accuracy and dimension reduction rate.