平行N-FINDR演算法的分析應用於遙測高維影像之端元萃取

Analysis of Parallel N-FINDR Algorithm Based on Endmember Extraction for High Dimension Remote Sensing Images

許晉輔(Chin-Ful Shu)

指導教授 ： 張陽郎

國立臺北科技大學/電資學院/電機工程系所/碩士(2010年)

摘要

衛星遙測影像可以大範圍的顯示各種地形的區域概況，藉以了解在各種不同地形地貌的分布情形，但是潛藏在這些地形地貌的異常物與相異物質卻難以被辨識出來。加上衛星感測器持續不斷的更新開發，造成光譜資訊日益龐大，更加深了偵測異常物與相異物質的困難性。因此如何快速、有效率的針對異常物(anomalies)偵測與相異物質區別，成為了一門重要的課題。
如何快速且有效率地對高光譜影像進行異常物偵測和相異物質區別是本論文所探討的。由於異常物或是相異物質的光譜特性與背景影像有著極大差異，但其面積相對背景影像來的較小，因此可透過N-FINDR演算法找尋和背景影像分佈不同的像素，來達到異常物偵測與相異物質區別的目的。然而其高資料量與複雜的計算相當費時，因此本論文透過使用『混合模型』(Mix-model)的技術來提高N-FINDR演算法對高光譜影像的異常物偵測與相異物質區別。
最後由實驗數據結果顯示，以平行N-FINDR演算法的Endmember萃取技術，能快速有效地針對影像做異常物偵測與相異物質分類，提升整體的運算效率。

關鍵字

Endmember ； N-FINDR ； CUDA ； GPU ； MPI ； OpenMP

並列摘要

This thesis is two twofold: (i) how to quickly and efficiently utilize the detection of hyperspectral anomalies, and (ii) to distinguish different substances. As what is well-known, the anomaly and the spectral properties of different substances depart from the background image in the tremendous degree. That is, its size is relatively smaller than the background images. One key solution lies in the method of the N-FINDR algorithm which function is to search for the pixel with different distributions from the background images. This method achieves the goals of the anomaly detection and the distinctions of substances.
However, the tasks of processing tremendously large data as well as the complex calculations are time-consuming problems which can be expected. As a result, this thesis adopts the technology of Mix-model to solve the techincal problems by expediting the N-FINDR algorithm to detect the image differences and the dissimilarities among materials of hyperspectral anomaly. The results of the experiments show that the Endmember extraction of parallel N-FINDR algorithm rapidly and efficiently achieves the hyperspectral anomaly detection. At the same time, the method also makes the fine-grained classification of different substances. Consequently, the overall operational efficiency is enhanced without a doubt.

並列關鍵字

Endmember ； N-FINDR ； CUDA ； GPU ； MPI ； OpenMP

參考文獻

[4] Winter, M. E.: N-finder: an algorithm for fast autonomous spectral endmember determination in hyperspectral data. In: Image Spectrometry V, Proceedings of SPIE, vol. 3753, pp. 266–277 (1999)

[6] S Sánchez ,Near real-time endmember extraction from remotely sensed hyperspectral data using NVidia GPUs , SPIE, Vol. 7724, pp. 772409-772409-9 (2010).

[7] Yang-Lang Chang, A Novel Approach to Hyperspectral Image Classification, Ph.D. Thesis, Chungli, Taoyuan, Taiwan , National Central University, May 2003

書籍：

[1] 張子浩著，「整合線性代數」，文笙書局出版。