物件式影像分析技術應用於土地覆蓋分類之研究

Land-cover Classification Using Object-based Image Analysis Approach

物件式影像分析（object-based image analysis, OBIA）為一種以物件式為影像分類基礎的新技術，近幾年已有潛力取代以像元為分類基礎的傳統式技術。本研究應用大地衛星（Landsat）影像搭配OBIA，進行恆春半島土地覆蓋分類之研究，首先，設定適當的尺度門檻，將影像分割為各自獨立的物件，再從各物件中挑選訓練樣區，作為最近相鄰演算法的基礎運算樣本，進而執行全區的影像分類，此外，也使用傳統式的最大概似分類法（maximum likelihood classification, MLC），比較兩種分類方法之成效。影像分類完成後，根據770個地面檢核點產生各土地覆蓋類型之混淆矩陣表並評估分類準確度，經評估結果發現，OBIA整體的準確度明顯優於MLC，且兩者之間具有顯著性的統計差異，綜合結果顯示，從衛星影像的觀測尺度，OBIA更有利於獲取更豐富地物特徵及空間分布格局，特別是在植被類型上的應用。

Object-based image analysis (OBIA) is increasing viewed as a more valid classification method than the traditional pixel-based method in recent years. In order to make comparisons of the two methods, in this study we applied OBIA with a Landsat image to implement land-cover classification in the Hengchun Peninsula of Taiwan. The OBIA approach involved segmentation of image data into objects at an agreeable scale level. Then objects were categorized using training set and the nearest neighbor algorithm. On the other hand, pixel-based classification was used in a maximum likelihood (MLC) algorithm. An accuracy assessment on both classifications using confusion matrices were undertaken based on 770 reference sites. A comparison of the results demonstrated a statistically significantly higher overall accuracy with OBIA over the MLC. The results suggest that OBIA can satisfactorily extract land information and explicit spatial patterns, and in particular vegetation from satellite imagery.