利用高光譜影像作異常物偵測

anomaly detection for hyperspectral imagery

陳獻廷(Hsien-Ting Chen)

指導教授 ： 任玄

國立中央大學/地球科學學院/太空科學研究所/碩士(2006年)

摘要

隨著遙測儀器技術的進步，高光譜影像在今日已被廣泛地使用。由於高光譜影像比多光譜影像擁有更好的光譜解析度，所以在目標物的偵測及物質分類方面的效果更好，相對的?了處理這些更龐大的資料，必須開發新的演算法來處理。本論文使用的是由I.S.Reed 和X.Yu所開發的RX演算法。RX演算法目的是在只有影像而沒有任何其他相關資訊的情況下，尋找影像中的異常物(anomaly)，這裡的異常物包含兩種特性：1.在整張影像中佔的面積極小；2.異常物的光譜值和背景物的光譜值有很大的不同。
RX演算法是先將資料做白化處理，再計算資料與原點間的歐基里德距離，但在異常物數目較多的情況下，RX演算法的效果明顯下降，因此本研究先利用模擬數據實驗對RX演算法的能力進行測試和分析，並提出ㄧ個利用PCA(principle component analysis，主成分分析)特性來預估背景物像素的方法，改善RX演算法在異常物數目增多時效果下降的問題。之後將此方法運用在AVIRIS和Hyperion高光譜影像上，結果顯示RX演算法在經過改良後，能完整的將異常物偵測出來，改善原先因異常物數目增多而偵測效果下降的缺點。

關鍵字

高光譜影像 ； 異常物偵測

並列摘要

With the improvement of remote sensing technology, hyperspectral imagery with higher spectrum resolution has uncovered many material substances which were previously unresolved by mutlispectral sensors. Anomaly detection has draw a lot of attention in hyperspectral image analysis recently. In general, such anomalous target are relatively small compared to the image, and their spectral signatures are distinct from their neighborhood. So it is difficult to detect anomalous targets especially with no prior information.
The RX algorithm was developed by Reed and Yu and assumes Gaussian noise and uses sample covariance matrix for data whitening. However, when the number of anomaly pixel exceeds certain percentage or the data is ill distributed, the sample covariance matrix can not represent the background distribution. In this case, the RX algorithm will not perform well. In this paper, we analyze the performance of the RX algorithm by computer simulation under different circumstances, including the number of anomaly pixels, number of anomaly types, the distance of anomaly spectrum from the background and the noise distribution. Then we improve the RX algorithm by utilizing characteristic of PCA (principle component analysis) to estimate the covariance matrix and mean of the pixels of the background and use two hyperspectral images (AVIRIS and Hyperion) to evaluate the performance. The experiment results prove that our method has better estimation of the background distribution and improves the performance of the RX algorithm.

並列關鍵字

anomaly detection

參考文獻

[1] G. Shaw and H. Burke, “Spectral imaging for remote sensing,” Lincoln laboratory journal, vol.14, no.1, pp.3-28, 2003.

[2] I.Reed and X. Yu, “Adaptive Multiple-Band CFAR Detection of an Optical Pattern with Unknown Spectral Distribution,” IEEE transactions on acoustics. speech. and signal processing, vol.38, no.10, pp. 1760-1770, 1990.

[3] C.-I Chang and S.-S Chiang, “Anomaly Detection and Classification for Hyperspectral Imagery,” IEEE transactions on geoscience and remote sensing, vol.40, no.6, pp. 1314-1325, 2002.

[4] H. Ren, Q. Du, and J. Jensen,“Efficient anomaly detection and discrimination for hyperspectral imagery,” Proc. SPIE, vol.4725, pp. 234-241, 2002.

[6] R. De Maesschalck, D. Jouan-Rimbaud, D. L. Massart, “The Mahalanobis distance,” Chemometrics and Intelligent Laboratory Systems, 2000