以高解析衛星影像輔以深度學習建置三維房屋模型;3D Building Model Reconstruction Using High Resolution Satellite Images with Deep Learning Analysis

NCUIR > Engineering College > Graduate Institute of Civil Engineering > Electronic Thesis & Dissertation > Item 987654321/88021

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/88021

Title:	以高解析衛星影像輔以深度學習建置三維房屋模型;3D Building Model Reconstruction Using High Resolution Satellite Images with Deep Learning Analysis
Authors:	呂振永;Lu, Chen-Yung
Contributors:	土木工程學系
Keywords:	高解析光學衛星影像;房屋平面圖;最小包絡矩形技術;正規化處理;約化處理;積木式三維房屋模型;High resolution optical satellite imagery;Building footprints;Minimum Bounding Rectangle (MBR);Regularization;Generalization;Block-based 3D building models
Date:	2022-01-20
Issue Date:	2022-07-13 15:14:26 (UTC+8)
Publisher:	國立中央大學
Abstract:	三維房屋模型可利用多種遙測技術所產生之資料進行建置，如衛星、無人飛行載具、航空測量及光達點雲等，而各方法皆有對於三維房屋模型建置的優缺點。近年來，衛星影像的空間解析度逐漸提高，已達數十公分等級，致使利用衛星影像建置三維房屋模型也逐漸受到討論。雖然衛星影像解析度與航測影像或光達點雲資料仍有差距，但衛星影像的涵蓋範圍廣、時間解析度高，因此對於建置三維房屋模型，仍有一定的優勢。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，從衛星影像自動萃取出房屋平面圖(Building Footprints)，並去除過小或不屬於房屋的多餘區域，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，塑形出較規律及方正的房屋多邊形。最後，匯入數值表面模型(Digital Surface Model, DSM)，在每棟房屋上層萃取附屬結構物，並利用其高程資料與RANSAC (RANdom SAmple Consensus)演算法擬合各多邊形高度，以建立積木式三維房屋模型。經過各項校正處理與去除異常值後，三維房屋模型平面(X、Y)誤差可達4.044公尺；高程(Z)誤差可達1.517公尺。由於衛星影像解析度較不足以建立複雜的屋頂模型，故建置模型之屋頂屬於平頂，精度符合OGC CityGML LOD1規範。後續三維房屋模型可應用於多種用途，如都市計畫、國土監測、災害重建、三維房屋模擬等，提供更接近真實世界的多元三維應用。 ;Three-dimensional building models can be reconstructed using data generated by a variety of remote sensing techniques, such as satellite, UAV, aerial survey, LiDAR point clouds, etc. Each method has advantages and disadvantages for the reconstruction of 3D building models. In recent years, the spatial resolution of satellite imagery has gradually improved, reaching tens of centimeters (cm) level, resulting in the use of satellite imagery to reconstruct 3D building models has gradually been discussed. Although there is still a gap between the spatial resolution of satellite imagery and aerial imagery or LiDAR point cloud data, satellite imagery has a wide coverage and high temporal resolution, so there are still 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 too small or not buildings. Next, Minimum Bounding Rectangle (MBR), regularization and generalization processing are utilized to shape the more regular and square building polygons. Finally, the Digital Surface Model (DSM) is incorporated to extract sub-structures on the upper floor of each building, 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 various corrections and the removal of outliers, the plane (X, Y) error and the elevation (Z) error of 3D building models can reach 4.044 m and 1.517 m respectively. Since the spatial resolution of satellite imagery is not enough to generate complex roof models, the roofs of reconstructed 3D building models is flat roofs, and the accuracy conforms to the OGC CityGML LOD1 specification. Subsequent 3D building models can be applied to a great deal of purposes, such as urban planning, land monitoring, disaster reconstruction, 3D building simulation, etc., providing a variety of 3D applications closer to the real world.
Appears in Collections:	[Graduate Institute of Civil Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	95	View/Open

社群 sharing

Loading...