外來物種的入侵，通常會對本土物種和生態造成影響。在南台灣的恆春地區，原產中南美洲的銀合歡，因過去人為不當的決策和其本身優越的生長能力，在當地大量的繁殖，威脅到當地豐富的生態。 目前對於外來植物的監測，大多採用人工現地調查，或是人工判釋航照、衛星影像的方式來進行，但成果精度掌握不易且不甚客觀。如採自動化分類的方式，只含少數波段的多光譜影像光譜資訊不足，常無法仔細且精確地區分植物，亦不能真正達到監測的目的。隨著遙測技術的進步，遙測影像的波段數從以往個位數波段的多光譜提升至上百個波段的高光譜，高光譜提供了大量且連續的高維度光譜資訊，因此，如何降低維度且萃取出有利植物間分類的資訊，為本研究重要課題。而PCA (Principle Component Analysis)主軸轉換為一般高光譜影像降低維度萃取特徵資訊的常用方法之一，不過其缺點為可能在特徵萃取時忽略差異較小但卻有助分類的資訊。因此，本研究使用在光譜上切割區塊的區段式PCA (Segmented Principle Component Analysis)，企圖更有效從高光譜影像中尋找銀合歡與其他植物間之差異，進而分類判釋銀合歡在恆春地區的覆蓋範圍。 本研究成果顯示，區段式PCA光譜轉換較能有效萃取出恆春地區銀合歡與其他植物之差異，分類成果較PCA佳。 Invasive species usually cause enormous ecological and environmental impacts and alter the ecological balance. In Heng-Chun area of southern Taiwan, Leucaena from South America spreads rapidly because of past human decisions and its flourishing adaptability. In order to make this problem under control and to develop strategies to maintain or restore local bio-diversity, it is necessary to understand the current status of Leucaena spreading. Traditionally, detecting invasive plants relies heavily on field investigations or human interpretations of aerial photos or satellite images. It is time-consuming and the result might not be reliable. On the other hand, the results of using automatic classification are often limited by multispectral data because there is not enough information for detecting specific plant in the limited spectral bands. To overcome this limitation, hyperspectral remote sensing data may be of more help. Hyperspectral remote sensing images provide more complete and detailed spectral information about ground coverage and have a great potential to the identification of specific plant species in vegetation covered areas. However the high data dimensionality of hyperspectral data can cause substantial impact to its applications. Principal component analysis is a common technique used for feature reduction in remote sensing image analysis, but it may overlook subtle but useful information. This research developed a segmented principal component analysis scheme that can be used not only to reduce the dimensionality of a hyperspectral image but also to extract critical spectral features helpful in discriminating different vegetation types. The analysis results of this research demonstrated that segmented principal component analysis performed better than regular PCA in providing critical information for distinguishing the target plant species from other vegetation types.
