| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 92 === In this study, the empirical mode decomposition (EMD) has shown its use in diverse applications for hyperspectral data analysis. Two issues have been concerned: dimensionality reduction and spectral unmixing.
In remote sensing, the number of labeled samples is usually limited, which makes the classification performance to be cursed by data dimensionality. To dispel the curse, some feature extraction techniques have been developed to reduce data dimensionality without loss of class separability. However, most of these techniques require accurately estimated class statistics in the full dimensional space in order to extract relevant features. Once the number of training samples is not large enough compared to data dimensionality, the estimates of statistics may not be adequately accurate for extracting appropriate features. It would be wise to perform a parameter-irrelevant processing to reduce dimensionality preliminarily for better estimated statistics.
A method is proposed in this study based on the empirical mode decomposition to construct a low dimensional space spanned by intrinsic mode functions (IMFs) for better estimation of class statistics. We have also discovered that IMFs are distributed in clusters. The empirical mode decomposition has recently been investigated in the blind source separation problem. Since it is analogous to spectral unmixing, we have applied the empirical mode decomposition for spectral unmixing.
The proposed approaches are tested on hyperspectral data containing various agricultural crops that could hardly be discriminated. Our experimental results demonstrate the feasibility of the EMD for preliminary dimensionality reduction and a potential use for spectral unmixing analysis. In dimensionality reduction, the proposed approach has given a relatively promising performance, compared to other methods such as the principal component analysis, the projection pursuit, and the wavelet approach. In spectral unmixing problem, the proposed approach has shown its potential in finding constituents of pixels.
|
|---|
