房屋區塊偵測為重建房屋模型與變遷偵測前之重要工作，最常見的資料來源之一為航照影像。航照影像可提供幾何、光譜與紋理資訊來達成偵測房屋區塊之目的。本研究利用高重疊數位航照影像，進行房屋區塊偵測，處理重點分為幾何分析、光譜分析、紋理分析與房屋區塊偵測四部分。幾何分析之主要工作為特徵主軸分析與多重影像匹配；光譜分析是藉由常態化差異植生指標(NDVI)或綠指標(GI)來分辨植生區；紋理分析是利用灰階共生矩陣(GLCM)中之角度二次矩法(ASM)來找尋均質區；最後，將幾何、光譜與紋理資訊整合，並利用最大相似法(ML)、Dempster-Shafer(DS)及支持向量機(SVM)三種不同的分類法，配合少量訓練資料以達到房屋區塊偵測之目的。 實驗成果顯示，多重影像匹配中特徵主軸配合提出的多視窗匹配方法，可有效提升匹配成功率與正確率。將房屋區塊偵測成果與參考資料相比，以像元為基礎之分類精度可達到80%且Kappa精度指標高於0.7；若以區域為基礎，由參考資料與成果套疊來看，皆可成功偵測到參考資料中房屋區之位置。Building detection is important in building reconstruction and change detection. The major information contents in building detection are shape, spectrum and textual when images are employed. This study uses highly overlapped aerial images to perform building detection. Geometry analysis, spectrum analysis, textual analysis and classifications are the important parts in this study. Geometry analysis includes two major works, which are multiple image matching and multiple feature direction analysis. On the other hand, the Normalized Difference Vegetation Index (NDVI) or Greenness Index (GI) is used to separate vegetation from buildings. Besides, textual information can find the homogeneity area using Angular Second Moment (ASM) based on Gray Level Co-occurrence Matrix (GLCM). Final, we employ Maximum Likelihood (ML), Dempster-Shafer (DS) and Support Vector Machines (SVM) to perform classification using integrated data sets. The results show that multiple image matching with feature direction analysis and the proposed matching strategy can improve the matching successful rate accuracy. Compared with reference data, the accuracy are higher than 80%, and Kappa index value about 0.7 in pixel-based validation. For region-based, the building regions of all the tests data with three different classifiers are detected successfully.
