利用衛星影像建置三維房屋模型逐漸受到討論,衛星影像的涵蓋範圍廣、時間解析度高,對於建置三維房屋模型,有一定的優勢。本研究主要針對以高解析光學衛星影像進行影像分析,並建立符合OGC CityGML LOD1等級之三維房屋模型。研究中應用深度學習,自動萃取出房屋平面圖(Building Footprints),並去除非房屋多餘區域,之後利用最小包絡矩形(Minimum Bounding Rectangle, MBR)技術、正規化(Regularization)及約化(Generalization)處理後,塑形出較規律房屋多邊形。最後,萃取房屋上層附屬結構物,並利用高程資料與RANSAC(RANdom SAmple Consensus)演算法擬合各多邊形高度,建立積木式三維房屋模型。經過校正與去除異常值,三維房屋模型平面及高程誤差皆可符合OGC CityGML LOD1規範。
The use of satellite imagery to reconstruct 3D building models has gradually been discussed. Satellite imagery has a wide coverage and high temporal resolution, so there are certain advantages for reconstructing 3D building models. This research focuses on image analysis based on high resolution optical satellite imagery and 3D building models reconstruction with an accuracy of the OGC CityGML LOD1 level. Deep learning technique is applied in this study to automatically extract building footprints from satellite images and remove excess areas that are not buildings. Next, Minimum Bounding Rectangle (MBR), regularization and generalization processing are utilized to shape the more regular and square building polygons. Finally, sub-structures on the upper floor of each building are extracted, and the elevation data is used to fit the height of each polygon with the RANSAC (RANdom SAmple Consensus) algorithm to reconstruct block-based 3D building models. After corrections and removing outliers, the accuracy of the reconstructed 3D building models conforms to the OGC CityGML LOD1 specification.