利用遙測影像探討東沙島周圍海草床分布

Abstract

本研究對2009年夏季的東沙島周圍海草床，以高解析度捷鳥（Quickbird）衛星影像進行海草範圍與不同海草群落的分布描繪，透過基本的兩種影像分類方法探討遙測技術應用於多物種組成之海草床的準確度，並嘗試加入海域深度資料以改善影像分類的結果，藉此產生全景、連續且尺度精細的海草分布地圖，同時探討東沙島周圍海草床可透過光譜區分出的海草群落組成，以及在光譜上易造成混淆的棲地資訊，並針對海草床內部特徵進行相關討論。 本研究以非監督式分類法繪製海草床範圍，產生整體準確度達76%的棲地分布圖，並找出可透過影像光譜區分出的海草群落，進一步將海草床區分成三種海草組成，在結合了非以海草為主要覆蓋物的海床棲地類型後，產生整體準確度達59%的海草群落與棲地分布圖；此外，本研究以監督式分類法對相同類別進行分析，在海草範圍上產生整體準確度為59%的分布圖，以及整體準確度為46%的海草群落與棲地分布圖。就海草類別的生產者準度上，非監督式分類可達97%，監督式分類可達63%；在海草類別的使用者準度上，非監督式分類為76%，監督式分類則可達84%。經由假設檢定發現在海草範圍的分布描繪上，非監督式分類法比監督式分類法擁有較高的分類準確度，然而不論是非監督式分類或監督式分類，深度資料的加入皆無顯著改善影像的分類準確度，可能是由於東沙島周圍海域深度淺且地形平緩，故在影像分類上並不用考量深度對遙測的影響。 研究結果中，從光譜分離度分析可發現，珊瑚的光譜特徵與有出現鋸齒葉水絲草（Cymodocea serrulata）或水韭菜（Syringodium isoetifolium）的海草群落相似；海草種類間的頻繁共存，亦使影像分類難以區分開各個種類的海草，如圓葉水絲草（Cymodocea rotundata）與泰來草（Thalassia hemprichii）因時常混生而難以透過光譜將兩者分離。 整體而言，對於多物種且混生情形複雜的東沙島海草床，影像分類技術的使用會有部份限制，但透過衛星影像繪製海草床或海草群落的分布圖可解決東沙島地處偏遠、交通不便的問題，對於海草床的監測與保育管理亦有相當大的幫助。

This study is to understand the capability of Quickbird satellite image for mapping the spatial distribution of seagrasses around the Dongsha Island, where the seagrass meadow there is composed of at least six species and occupies an area of 12 km2. Through two methods of image classification, both seagrass coverage and communities with different coexisted assemblages of seagrass species can be mapped, and the thematic maps produced can provide high-resolution and continuous distribution message. The results showed that using the unsupervised classification to map the seagrass coverage, the thematic map had an overall accuracy of 76%, as well as the producer’s accuracy of 97% and the user’s accuracy of 76% for the seagrass class. When the seagrass meadow was divided into three types of seagrass communities by the unsupervised classification, the overall accuracy was 59%. By using supervised classification, the thematic map had a lower overall accuracy on the seagrass coverage and communities (59% and 46%, respectively), while the producer’s and user’s accuracy of the seagrass classes combined were 63% and 84%. The accuracy of the unsupervised classification was significantly higher than that of the supervised classification in mapping the seagrass coverage. No matter which classification was used, the input of the depth data didn’t improve the accuracy of the image classification, which might be owing to the shallowness of the water and the gentleness of the terrain around the Dongsha Island. Moreover, this study found that some compositions of seagrass species had similar spectral features with some other habitats, which may cause misclassification. For example, the spectral separability between corals and the seagrass communities with Cymodocea serrulata (or Syringodium isoetifolium) was poor. As the frequent mixture of seagrass species also led to the difficulty of spectral discrimination between some species pairs, like Cymodocea rotundata and Thalassia hemprichii. Although remote sensing techniques have some limitations in complex environments, such as image classification in the Dongsha Island, it is still an cost-effective way to monitor seagrass meadows located in faraway places, and is quite helpful to seagrass conservation and management.