『高光譜』遙測影像 (Hyperspectral Imagery) 現為遙測影像最新之技術，遙測影像頻譜解析度由原先數個頻譜解析度的一般感測器 (如SPOT 5)、至數十個頻譜解析度之『多頻譜感測器』 (Multispectral)、到數百個頻譜解析度的『高光譜感測器』 (Hyperspectral)、乃至於數千個頻譜解析度之『超高光譜感測器』(Ultraspectral)， 此一『高光譜』解析度之感測器已廣泛地應用於衛星遙測影像之識別、醫學影像的診斷檢查、工業產品之檢驗、飛機及其他精密機器設備之非破害性檢查等之應用，此一領域之研究，正如火如荼於全球各地、方興未艾熱烈地擴展當中。 我們提出一新方法適用於『高光譜』遙測影像分類，其中有兩個主要的實現方法，第一個方法為『貪婪模組特徵空間』(Greedy Modular Eigenspaces)，第二為『布林濾波器』(Positive Boolean Function)。並藉由實際校正過後的美國國家航空太空總署(NASA)所提供之完整台灣『高光譜』遙測影像資料，以及國立中央大學太空及遙測研究中心所實地測量的台灣地表真實資料，來驗證『貪婪模組特徵空間』的方法確實提供了一個絕佳的特徵抽取方式，並成為一個最適合『布林濾波器』分類方法的前處理器。 本論文將詳細討論『貪婪模組特徵空間』新方法理論之推導、提供完整的『布林濾波器』基礎理論依據，及詳細分析『貪婪模組特徵空間』與『布林濾波器』之關係，並針對二者的特性關係，加以推演修正後，進而推廣並提出解決『高光譜』與『合成孔徑雷達』影像資料融合的問題方法。最後經由所設計之驗證實驗，實際操作於台灣『高光譜』遙測影像資料上，並將之與其他傳統應用於多頻譜感測器遙測影像資料分類方法作一效能之比較，印證了本方法非常適用於『高維資料』(High-Dimensional Data)分類的特性。 Tremendous efforts have been focused on the developing of hyperspectral imagery classifications devoted to earth remote sensing. This dissertation presents a new supervised classification approach to hyperspectral imagery, which consists of two algorithms, referred to as greedy modular eigenspace (GME) and positive Boolean function (PBF). We first introduce a GME, which is a modification of the original module of a complete modular eigenspace (CME), obtained by a quick band reordering greedy modular eigenspace transformation (GMET) algorithm. The proposed GMET algorithm is very efficient with little computational complexity. A GME can be treated as not only a preprocess of the PBF-based classifier but also a unique feature extractor to generate a particular feature eigenspace for each of the material classes present in hyperspectral data. The features extracted from hyperspectral images by this algorithm are proven by our experiments to be crucial for the subsequent PBF-based classification. The GME makes use of the potential significant separability of different classes embedded in the correlation of hyperspectral data sets to overcome the drawback of the common covariance pool bias problems encountered in conventional principal components analysis (PCA). It uses the data correlation matrix to reorder spectral bands from which a group of feature eigenspaces can be generated to reduce the dimensionality. The residual reconstruction errors (RRE) are then calculated by projecting the samples into different individual GME-generated modular eigenspaces. The PBF is a stack filter built by using the binary RRE as classifier parameters for supervised training. It implements the minimum classification error (MCE) as a criterion so as to improve the classification performance. It utilizes the positive and negative sample learning ability of MCE criteria to improve the classification accuracy particularly in dealing with hyperspectral data in which training data are always inadequate. The proposed PBF-based classification scheme is developed to effectively find nonlinear boundaries of pattern classes in hyperspectral data. It possesses well-known threshold decomposition and stacking properties. The advantage of PBF-based classifiers are their truly exhaustive discrete and nonlinear binary properties. This characteristic can best harmonize the PBF-based classifiers with the features extracted from GME. It improves classification accuracy extraordinarily and fully promotes multi-classifiers instead of pairwise-classifiers. The combining of the GMET algorithm with the PBF-based classifier provides a tremendously unique advantage to hyperspectral image classification. Moreover, high-dimensional spectral imageries obtained from multispectral, hyperspectral or even ultraspectral bands generally provide complementary characteristics and analyzable information. Synthesis of these data sets into a composite image containing such complementary attributes in accurate registration and congruence would provide truly connected information about land covers for the remote sensing community. In this dissertation, we also propose a novel feature selection algorithm applied to the GME to explore a data fusion technique using data fused from data gathered by the MODIS/ASTER airborne simulator (MASTER) and the Airborne Synthetic Aperture Radar (AIRSAR). The proposed approach, based on a synergistic use of these fused data, represents an effective and flexible utility for land cover classifications in earth remote sensing. The proposed GME method has the advantage of preserving the individual abundant features in different classes and, as far as possible, avoiding dependence on global bias statistics. GME significantly improves the precision of image classification compared with conventional feature extraction schemes. Experimental results demonstrate that the proposed GME feature extractor suits the nonlinear PBF-based multi-class classifier perfectly well for classification preprocessing. Compared to the conventional PCA, it not only dramatically improves the eigen-decomposition computational complexity but also consequently increases the accuracy of image classification. Experiments also show that the Vague boundary sampling properties can make the process of labeled sample selection from hyperspectral data more practicable and efficient. These remarkable features will be presented in this dissertation.
