三維模擬退火波段選擇方法之布林函數分類器研究

A Positive Boolean Functions Classifer Based on Three-dimension Simulated Annealing Band Selection Approach to Hyperspectral Images

賴嘉琪(Jia-Chi Lia)

指導教授 ： 梁文耀 ； 張陽郎

國立臺北科技大學/機電學院/通訊與資訊產業研發碩士專班/碩士(2008年)

摘要

衛星遙測影像可以展示大範圍區域的地形概況，藉以了解各種不同地形地貌，如山脈、平原、盆地、海岸線、城市、河流、道路等的分布狀況；更細緻的地形分布，如山勢的走向、河流的流向與幾何形貌等，也可以一覽無遺。而在遙測影像中，衛星感測器持續不斷的更新開發，造成可利用的光譜資訊日益龐大，而其中包含雜訊或是錯誤資訊的光譜不在少數，因此挑選正確有效的光譜資訊進行分類，特徵波段選取的前置處理就愈顯得重要。
而本論文採用一個基於三維的模擬退火法特徵抽取技術，有效地進行高光譜特徵抽取再行分類，本篇論文分別以四種分類器:布林函數分類器(Positive Boolean Function, PBF)、最大近似分類器(Maximum Likelihood, ML)、支援向量機分類器(Support Vector Machine, SVM)、k近鄰分類器(k Nearest Neighborhood, kNN)進行研究分析。PBF分類器以改善樣本比對的方式，加強分類的正確率；ML是依據機率(Probability)的概念所產生的分類器；SVM分類器使用核心函數，進行訓練學習；而kNN則利用與實例樣本之間的距離，進行分類動作。由實驗數據結果顯示，本論文提出的以三維的模擬退火法波段選取方法針對PBF分類器的正確率明顯比其他的分類器來得高，並且能夠作為現今分類器演算法中的另一種選擇。

關鍵字

布林函數分類器(Positive Boolean Function) ； 最大近似分類器(Maximum Likelihood) ； 支援向量機分類器(Support Vector Machine) ； k近鄰分類器(k Nearest Neighborhood) ； Classifier

並列摘要

A sophisticated remote sensing technique was introduced for decades to display and realize various topographical features, such as the distribution of mountains, plains, basins, coastlines, cities, rivers and roads, over a large area of terrain. Moreover, a more detailed topographic distribution, such as the direction of the mountain, river flows, geometry morphology, can also be examined by this modern satellite technology. In telemetry images science, the technologies of satellite sensors are increasingly developed, resulting in the huge availability of spectrum; however, minor noise and erroneous information are also detected. Therefore the selection of the correct and effective spectral information to be classified, and pre-processing of the characteristics sampling are apparently significant.
Thesis adopts the simulated annealing method of three-dimensional (3D) feature extraction technology to re-classify the Hyperspectral feature extraction by four classifiers: Positive Boolean Function (PBF), Maximum Likelihood (ML), Support Vector Machine (SVM), and k Nearest Neighborhood classifiers (kNN). PBF classifier improves the sampling ratio, and strengthens the correction. ML classifier is generated based on the probability; SVM classifier uses the core function to implement learning training. kNN classifier utilizes the distance between examples and itself to classify. This thesis proposes a 3D simulated annealing (3DSA) feature extraction approach for the band selections prior to the PBF classifications. Based on the experimental results, the proposed 3DSA/PBF method presents a higher accuracy compared to other classifiers. It can also be an alternative of the satellite remote sensing image classifications.

並列關鍵字

Positive Boolean Function (PBF) ； Maximum Likelihood (ML) ； Support Vector Machine (SVM) ； K Nearest Neighborhood (kNN) ； Classifier

參考文獻

[19] 張陽郎、韓欽銓、范國清、陳錕山，「一個新穎的方法實現高光譜監督式遙測影像分類」，航測及遙測學刊， 9卷4期， 2004年12月。

[21] 林均傑， 一個對於高光譜影像的二維模擬退火波段選擇方法，碩士論文，國立台北科技大學，台北，2007年6月。

[23] 陳木松、廖鴻翰，「適應性突變運算及其應用」，大葉學報，7卷1期，1998年，pp 91-101.

[1] Michalewicz, Z., “Genetic Algorithms + Data Structures = Evolution Programs”,

[2] Kirkpatrick, S., Gelatt, C.D., and Vecchi, M.P., “Optimization by Simulated Annealing,” Science, Vol. 220, pp. 671-680, 1983.