本論文在嘗試結合非監督型類神經網路（unsupervised artificial neural network）與聚類分析（clustering analysis）對衛星影像作快速分割（segmentation）及分類（classification）。其過程分為三階段：第一階段，利用概括式編碼方法對遙測影像之光譜資訊進行編碼，第二階段利用可調適性共振理論第二型（ART2）類神經網路對編碼後的光譜資訊進行分類以獲取頻譜類別（spectral classes）〔Ric86〕，而第三階段，運用第二階段中所獲得的頻譜類別，藉由fuzzy class C - means聚類演算法重組得個數上較少的資訊類別（information classes）〔Ric 86〕。由於fuzzy class C - means演算法，必須告知要聚類的數目。為了要達到自動化的目的，設計一種能描述聚類結果好壞的指標函數。經由此函數，決定出最佳的聚類數目。以上所提出的技術將應用在人造衛星影像與真實位星影像上。
研究結果發現，雖然沒有比對的資料以辨別真實影像分類結果的好壞，但是從顯示出的結果中可發現，其辨識率與監督式分類器不相上下，而且沒有了監督式分類器，如依賴使用者的操縱、資料取樣的好壞以及費時等限制與缺點。

An unsuperivsed classification approach conceptualized in terms of neural and fuzzy disciplines for segmentation of remotely sensed images is presented. The process consists of three major steps: data transformation, neural classification, and fuzzy grouping. In the first step, multispectral patterns of image pixels are transformed into what we call coarse pattems using a histogram - based nonuniform coarse coding technique. Coarse patterns not only elimimate a proportion ambiguity existing amog spectral patterns but increase their separability. In the second step, a fine classification of image pixels is achievd by applying an ART neural classifier to the coarse pattems of pixels. The resultant clusters of pixels are called spectral classes. Since spectral classes are typically too detailed to be of practical signicance, in the third step a modified fuzzy c - means algorithm, called the fuzzy class c - means algorithm, is invoked to integrate spectral classes of pixels into informative classes. A function for evaluating the goodness of classification is defined by which the optimal number of classes within a given range can be automatically determined. The proposed technique is applied to both synthetic and real (SPOT satellite) images. Experimental results are shown which support the applicability of the proposed method, and its performance can be comparable to those of supervised ones yet without their infirmities such as user involement, time consumption, and data dependency.
Index terms - Spectral classes, Informative classes, Histogram - based nonuniform coarse coding, Adaptive representation, ART neural classifier, Fuzzy clustering algorithm, Performance measure.