比較單體基底的高光譜影像端元萃取方法

Comparison of Simplex-Based Endmember Extraction Algorithms for Hyperspectral Image

吳至誠(Jee-cheng Wu) ； 郭純吟(Chun-yin Guo) ； 李國雲(Kuo-yun Li)

《國立宜蘭大學工程學刊》 10期 (2014/12) Pp. 93-104

https://doi.org/10.6176/BCE.2014.10.17

摘要

雖然高光譜影像具更大量的資訊，但影像受空間解析度與地物分布之影響，它存在著混合像元。本文僅針對兩種主要的單體基底（simplex-based）端元萃取方法（內部最大單體法與最小體積包圍單體法），於高光譜影像中萃取純端元（單一物質），並評估其性能。首先參考地面真實地圖於高光譜影像中選取10種不同種類的端元；其次於高光譜影像中以兩種單體基底端元萃取方法萃取出七種不同端元位置；接著以完全約束最小二乘法將萃取出之端元製作成不同端元之豐度圖，並比較各豐度圖與地面地質圖以確認各端元名稱；最後以光譜角製圖計算出各萃取出之端元與地面參考地物間之夾角，並分析兩種單體基底端元萃取方法之效能。實驗結果顯示內部最大單體法比較能找到較多不同類別的端元，但其計算非常費時。

關鍵字

高光譜影像 ； 光譜分解 ； 端元萃取

並列摘要

In hyperspectral image, there exist mixed pixels because of the image spatial resolution and ground object distribution. In this paper, we focus on only two major simplex-based endmember extraction methods, i.e. N-findr (or N-FINDR) and Minimum Volume Enclosing Simplex (MVES). We aim at extracting endmember (single material) from hyperspectral image and assessing the performance of the two methods with the following four steps. Firstly ten endmembers from the ground geological map are selected. Then, the two simplex-based endmember extraction methods are used to find out the location of the seven endmembers in hyperspectral image. Thirdly the extracted endmembers using fully constrained least squares method are applied to generate abundance maps. Besides, the abundance maps are compared with the ground geological map to confirm each endmember of the name. Finally we compute spectral angle mapper (SAM) between the extracted endmember and the ground reference endmember to evaluate the performance of the two simplex-based endmember extraction methods. The experimental results showed that more endmembers can be extracted with N-FINDR than with MVES, but N-FINDR calculation is very time-consuming.

並列關鍵字

Hyperspectral image ； Spectral unmixing ； Endmember extraction