| 摘要: | 台灣的牡蠣養殖業具有悠久的歷史，並且台灣的環境和氣候非常適合牡蠣的 全年繁殖。主要的養殖地區位於雲林、嘉義和台南的沿海地區，其中雲林縣是主 要的牡蠣苗生產區。牡蠣養殖方法根據不同水深可以分成浮筏式棚架(浮棚)與 垂下式棚架(立棚)。牡蠣產量的減少對雲林當地的蚵農造成了生計上的困難， 然而，由於土地劃分和統計資料真實性等問題，對於雲林沿海地區的牡蠣養殖業 缺乏完整且可靠的資訊。因此，本研究旨在針對雲林縣台西鄉的牡蠣養殖範圍進 行提取，利用多時期的 SPOT 衛星影像辨識牡蠣棚架，以獲取歷年的牡蠣養殖範 圍和棚架數量。傳統的像元分類方法對於海洋環境中的養殖範圍提取具有一定的 挑戰，因為複雜的海水背景和光譜訊息相似的養殖範圍會影響分類結果。物件式 影像分析(object-based image analysis, OBIA)方法可以解決這些問題。
本研究使用 OBIA 方法對 SPOT 影像進行分類，並選擇了支持向量機(Support Vector Machine, SVM)作為分類器進行兩次試驗，對空間解析度 6 公尺之多光譜 影像，以及其 1.5 公尺之銳化影像進行分類，同時結合光譜(標準化植生指數、 標準化水體指數、明亮度、光譜平均值、光譜標準差)和 GLCM(Gray-Level Co- Occurrence Matrix)紋理特徵(能量、熵、相關性、逆差矩)等作為分類的依據。 試驗一針對整張影像進行分類、試驗二則是將影像分成浮棚與立棚兩個區域再進 行分區分類，也同時比較僅使用光譜特徵以及同時使用光譜與紋理特徵之分類精 度。結果顯示，試驗一中所有分類結果整體精度皆高於 85%，表示 OBIA 方法能 夠在 SPOT 影像中準確識別牡蠣棚架，且當分類條件中加入能量與逆差矩紋理特 徵時，能夠有效提高分類的精度。而在試驗二分區分類結果中，浮棚的整體精度 皆高於 90%，而立棚在 6m 影像中加入紋理特徵的整體精度為 65.49%、1.5m 影 像中則是 82.3%，由此可知浮棚的分類精度比立棚高，且使用高解析度影像可以 得到更好的牡蠣棚架分類結果。
另外本研究獲取了 2003 年至 2020 年間雲林縣台西鄉的牡蠣養殖範圍和棚 架數量。由結果得知過去主要以立棚為主、分佈範圍廣泛，但因為地形變化，2010 年後立棚逐漸減少、浮棚漸增。整體而言過去 18 年來台西鄉在採苗時期的牡蠣 棚架逐漸減少。這項研究有助於填補雲林沿海地區牡蠣養殖業資訊不完整的缺 陷，並提供了重要的參考依據，以支持牡蠣養殖業的發展和管理。;Taiwan′s oyster aquaculture industry has a long history, primarily in Yunlin County for oyster seed production. Oyster farming methods include floating and overturned frames, depending on water depths. However, declining oyster production challenges local farmers′ livelihoods. Limited by land division and data reliability issues, there is incomplete information about oyster aquaculture in Yunlin′s coastal areas. This study uses multi-temporal SPOT satellite imagery and object-based image analysis (OBIA) to identify oyster cultivation areas in Yunlin′s Taixi Township. The Support Vector Machine (SVM) was employed as the classifier in two experiments using SPOT images with spatial resolutions of 6 m and 1.5 m. Spectral features including Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), brightness, spectral mean, and standard deviation, and GLCM (Gray-Level Co-Occurrence Matrix) texture features, including energy, entropy, correlation, and inverse difference moment were used in the classification. In Experiment 1, the entire image was classified, while in Experiment 2, the image was divided into floating frames and overturned frames. Meanwhile, the SVM was conducted and compared when only spectral features were used and both spectral and texture features were used. The results of Experiment 1 show that the overall accuracies of classification results are all above 85%, indicating that the OBIA method accurately recognizes oyster beds in SPOT images. Furthermore, the inclusion of texture features of energy and inverse difference moment effectively improves classification accuracy. In Experiment 2, with combining spectral and texture features, the classification accuracies of floating frames are consistently higher than 90%, while the classification of overturned frames has overall accuracies of 65.49% and 82.3% when applying 6 m and 1.5 m images, respectively. This indicates that the classification accuracy of floating frames is higher, and images with higher spatial resolution result in better oyster frames classification. Additionally, the study mapps oyster cultivation areas and frame numbers from 2003 to 2020 in Taixi Township. Overturned frames were common in the past but gradually decreased after 2010, while floating frames increased. In the past 18 years, the number of oyster frames during the larvae period declined in Taixi Township. This study bridges the information gap in Yunlin′s coastal oyster aquaculture, providing essential references for development and management. |
