基於粒子群優演算法的多屬性決策-非純度波段優先權方法應用於高維度資料波段選取

Particle Swarm Optimization-based Impurity Function Band Prioritization Using Multiple Attribute Decision Making Model for Band Selection of High Dimensional Data Sets

林國凱(Kuo-Kai Lin)

指導教授 ： 張陽郎

共同指導教授 ： 方志鵬

國立臺北科技大學/電資學院/電機工程系所/碩士(2014年)

https://doi.org/10.6841/NTUT.2014.00295

摘要

近年來由於衛星遙測技術日益進步，衛星遙測影像的維度及資料量也日趨龐大，因此為解決衛星遙測影像在使用上之維度與資料量過於龐大複雜的問題，可利用波段選取來降低影像的維度，進而避免由於波段數增加所導致的Hughes現象。
先前有學者提出許多波段選取演算法來達到降維效果，然而這些波段選取演算法成效不彰，使得降維效果並不明顯。因此，本論文提出「粒子群優法」結合「相關係數矩陣」分別聚合各類別中的高相關度波段以獲得特徵模組空間，接著以「多屬性決策之層次分析法模型」將各類別的模組特徵空間與區塊彼此關係以層級化呈現，再結合「非純度波段優先權法之類別覆蓋率」計算各群聚波段中各波段之權重分數，最後將所有波段的加權分數進行統計，挑出代表性的波段，以達到降維的效果。
在實驗結果，本論文採用的遙測影像實驗圖資為MASTER的鰲鼓溼地及 AVIRIS的Northwest Tippecanoe County (NTC)，並測試不同的降維率與正確率之間的變化與關係；其中當鰲鼓溼地的降維率達到90.91%時，正確率為98.44%；NTC的降維率達到85.00%時，正確率為94.48%。

關鍵字

粒子群優法 ； 高維度資料 ； 波段優先權 ； 波段選取 ； 相關係數矩陣 ； 多準則決策 ； 層級分析法

並列摘要

In recent years, with the progress in remote sensing technologies, the numbers of data and dimension are increased in remote sensing imagery. For solve the huge data and high dimension computational complexity problem in remote sensing imagery can use band selection methods to reduced dimension and avoid Hughes Phenomena because of increased quantity of bands.
Some scholars proposed many kinds of algorithms for band selection to reduce dimensionality, but those algorithms were inefficient and led to the dimensionality reduction effects couldn’t be significantly. Therefore, in this paper, using a band selection algorithm based on particle swarm optimization (PSO) combine with correlation coefficients matrix (C.C Matrix) to cluster the highly correlated bands together and to obtain greedy modular eigenspace (GME) and represented as analytic hierarchy process (AHP) module to observe the relationship with modular hierarchically. To combine AHP with impurity function class overlapping (IFCO) to calculate high correlation bands weights. Finally, according to the bands weighting scores statistics result to select representative bands for obtaining greatly dimensionality reduction effect.
The effectiveness of the proposed method is evaluated by MASTER and AVIRIS remote sensing images for testing the variance and correlation of dimension reduction rate and classification accuracy. The experimental results show the dimension reduction rate is 90.91% and classification accuracy is 98.44% in Au-Ku; the dimension reduction rate is 85.00% and classification accuracy is 96.48% in NTC.

並列關鍵字

Particle Swarm Optimization (PSO) ； High Dimensional Data Sets ； Band Prioritization ； Band Selection ； Correlation Coefficient Matrix ； Multi Criteria Decision Making (MCDM) ； Analytic Hierarchy Process(AHP)

參考文獻

[17] 王旭，基於GPU之粒子群優法應用於高光譜影像波段選取，碩士論文，國立臺北科技大學電機工程研究所，臺北，2012。

[16] 徐斌峰，一個維度優先權方法應用於粒子群優法在高維度影像特徵抽取，碩士論文，國立臺北科技大學電機工程研究所，臺北，2011。

[18] 王品皓，基於粒子群優演算法的多數決非純度波段優先權方法應用於高維度資料特徵抽取，碩士論文，國立臺北科技大學電機工程研究所，臺北，2013。

[15] 王宏原，平行粒子群優法應用於高維度影像特徵抽取，碩士論文，國立臺北科技大學電機工程研究所，臺北，2010。

[5] Yang-Lang Chang, Chin-Chuan Han, Bormin Huang, Wen-Yen Chang, Benediktsson and J.A.,Lena Chang, "A Parallel Simulated Annealing Approach to Band Selection for High-Dimensional Remote Sensing Images," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol. 4, 2011, pp. 579 - 590 .