| 摘要: | 建置三維房屋模型的目的，在於提供真實物空間三維資訊與決策支援，如都市規劃更新、災害應變、建設管理等。由於房屋模型的建立仍未完全自動化，且房屋改變迅速，若進行變遷偵測，針對有變遷的房屋區域進行重新建置，將可減少資料更新所需的成本且提升效率。因此，為能有效更新，房屋模型之變遷偵測是重要的研究課題。傳統上常利用多時期影像之光譜差異進行變遷偵測，此方法僅有二維光譜資訊而缺乏三維形狀資訊。隨著光達系統成熟，使三維形狀資訊取得容易。本研究使用後期光達點雲及航照影像進行前期房屋模型之變遷偵測。 主要工作項目包含：(1)資料前處理，(2)判定原房屋模型變遷型態，(3)偵測新建與有變遷之房屋，(4)產生後期房屋區域。資料前處理，進行資料之套合，及剔除地面與植生區域之光達點雲，且計算前後期高程差異。判定原房屋模型變遷型態，結合光譜及形狀資訊進行判定；本研究設定五種變遷型態，分別為未改變、主結構改變、副結構改變、拆除、及植生遮蔽。第三步驟剔除非房屋光達點，找出新建與有變遷之房屋點雲。最後，利用未改變的區域及新建與有變遷之光達點產生後期房屋區域。 本研究成果於判定原房屋模型變遷型態部分，可達85%整體精度。產生之後期房屋區域以像元為單位之驗證顯示整體精度為96%，以區塊為單位之驗證顯示誤授率為4%及漏授率為13%。並且，為詳細了解影響研究成果的因素，本研究亦針對所有錯誤例進行分析。 3D building model provides spatial information for city planning, construction, and management. Because the reconstruction of building models is still not fully automatic and the cities change rapidly, it would be more preferable to maintain a building database that firstly detect the changes followed by a reconstruction procedure. Therefore, change detection of building model is an important issue for efficiently updating. Traditionally, change detection is usually done using multi-temporal images through the spectral analyses. Those images provide two-dimensional spectral information without including shape in the third dimension. As the availability and quality of emerging LIDAR systems that make the acquisition of shape information convenient, we use new LIDAR point clouds and aerial photos to detect changes for old building model. The proposed scheme comprises four major parts: (1) data pre-processing, (2) detecting changes on old building areas, (3) finding new or changed buildings, and (4) generation of new building regions. The first step performs the spatial registration for the different types of data. In addition, we remove LIDAR points in ground and vegetation areas. In the second step, we integrate shape and spectral information to determine the change type of building models. We set five change types in this research, namely, unchanged, main-structure changed, micro-structure changed, demolished, and vegetation occluded. In the third step, we search for new or changed buildings by removing non-building points. Finally, we use unchanged building regions and new or changed building points to derive new building regions. The validation for determination of change type shows that the results can reach 85% overall accuracy. The results for new building regions reach 96% overall accuracy by pixel-based validation. In region-based validation, the commission and omission errors are 4% and 13%, respectively. To provide comprehensive observations, those unreliable results are scrutinized. |
