2010-06-012024-05-15https://scholars.lib.ntu.edu.tw/handle/123456789/663946摘要：隨著衛星感測技術之進步，遙測衛星已邁向高空間及高光譜解析度之世代，其中高空間解析度衛星影像所呈現之地物類別遠比一般中解析度衛星影像豐富且複雜。傳統遙測影像分類方法大多採取逐像元分類方法，以單一像元之光譜反射值為分類依據，忽略掉鄰近影像區塊之紋理及幾何結構等資訊，若將此類方法應用於高解析度衛星影像分類時，將會造成嚴重的椒鹽效應(Salt and Pepper Effect)，且無論是在執行效率或分類準確性上，皆已無法滿足實務工作之需求。為提升分類效率，並避免分類錯誤現象，本研究將採區塊式之物件分類方法(Object-base Classification)，即先對影像進行分割(Image Segmentation)後得到各式各樣的區塊(Region)，一般稱之為物件，之後再計算各區塊物件之特徵(Feature)，用來描述該物件之屬性，作為影像區塊分類之依據。本研究將依據特定判釋目標或分類區域之特性，嘗試應用物件式分類方式，改進傳統分類演算法準確度與完整度不足之缺點，同時探討不同地貌下兩種分類方法之優缺點，並經由比較物件導向與傳統分類方法於執行效率及辨識準確度之差異狀況。此外，一般的物件導向分類方法中，可將物件特徵簡單區分為形狀、紋理、色彩等三類，目前針對不同應用所提出之物件特徵已達數百種，本計畫亦將研究物件導向分類方法於特定目標(如儲油槽、特定建築物、機場與陣地等)點位分類時，應採用何種空間屬性指標或特徵(例如：面積、週長、形狀、紋理等)，作為實務分類工作時之參考，逐步透過特定目標區特徵萃取，建立空間向量基資，以輔助影像分析、空間分析及變遷分析等情研工作。<br> Abstract: With the development of remotely sensors, high spatial and spectral resolution of satellite images have the ability to provide richer and more complicated information of earth surface than the moderate resolution satellite images. Most of the classification methods for satellite images are based on the spectral variations of a single pixel. This will cause the salt and pepper effect because the texture information and geometry structure in the neighborhood of the classified pixel are ignored. In order to avoid the salt and pepper effect and improve the classification efficiency, a new classification method based on the object-oriented theory is used in this study. The process of object-oriented classification can be divided into three simple steps. Firstly, objects are created during the segmentation process that the image is subdivided into groups of pixels with the similar local contrast value. Then different kinds of object features are calculated for the classification process. In this study, the differences between traditional and objected-oriented method are firstly compared, and the efficiency and accurate assessment are also evaluated. In addition, object features can be derived from the shape, texture, and spectral (color) information of objects. Up to now, hundreds of features have been proposed for different applications. In this study, the useful features for some specific surface objects will be identified and used for classification. Finally, the classification results will be imported into the database for future applications of image analysis, objected indentation and change detections.影像分割影像分類物件導向Image SegmentationImage ClassificationObject-Oriented