以無人飛行載具影像技術於水利建造物之巡檢研究

Abstract

水利建造物巡檢為確保水利設施安全的例行工作，現行作業方式主要採人力現地調查，耗費大量人力及時間成本，且危險性高，而地形的障礙是導致現場檢查效率低的主因，管理機關為克服地形障礙引入無人飛行載具(Unmanned Aerial Vehicle, UAV)拍攝空拍影像輔助檢查，現行做法單張空拍影像缺乏空間資訊，檢查人員不易判讀檢測位置及量化劣化範圍，不利於評估劣化嚴重性。本研究流程主要分成五大部分，首先影像蒐集與預處理，包含目標區參考底圖及巡檢影像拍攝、目標區數值地形模型、正射影像產製；再單張巡檢影像相機外方位重建，將巡檢影像與參考底圖進行特徵點匹配，萃取地面控制點，以單張空間後方交會平差解算外方位；接著影像缺失判讀，由於研究上限制，以植生偵測模擬劣化樣態做為範例，分別以機器學習方法及遙測指標，辨識水利建造物及植生範圍，計算劣化樣態在水利建造物上比例，作為劣化嚴重性指標；然後巡檢影像幾何校正及劣化區域定量，將多時期巡檢影像根據參考底圖數值地形模型進行正射化糾正，針對非正攝區域則採用影像建模與劣化區定量。最後成果分析與品質評估，對於外方位解算成果、劣化區定量及影像缺失判讀進行正確性驗證。本研究蒐集大安溪義理橋河段水利造物影像，經過幾何校正的巡檢影像，其定位精度可達公分級。影像缺失判讀部分，透過深度學習模型進行水利建造物分類中及遙測指標進行植生分類，結果顯示準確度以達到90%以上，說名影像缺失判讀程序可行，未來可以擴展至各種劣化樣態的偵測。本研究基於UAV影像技術易於執行的標準化作業流程，並從影像中獲取足夠空間資訊以輔助判斷建造物安全性，發展可以高效率且準確定性定量的技術，提升水利造物調查效率，在河川防災及整治中能夠有所助益。

Hydraulic structure facility requires routine inspection to ensure the safety of the structure. The traditional inspection methods are mainly on-site investigation, which costs a considerable amount of time and human resources and poses high safety risk. The major factor that affects the efficiency of the inspection is the terrain where the structure resided. To overcome this obstacle, Unmanned Aerial Vehicle (UAV) was introduced, which can take aerial images remotely to assist the task. However, pure aerial images often lack sufficient spatial information and thus make the inspector unable to precisely identify the position, the size as well as the area of the deterioration on the hydraulic structure structures, increasing the difficulty in documenting the severity of the deterioration. This research is divided in five parts: first, image collection and pre-processing, including reference base map definition, inspection image acquisition using UAV and the production of Digital Surface Model (DSM) and orthoimage of the target area; next, the computation of the camera’s exterior orientation parameters. By performing feature points matching between the inspection image and the reference base map, one can extracts the ground control points, and solve for the parameters through Single Photo Resection. Third, we employ both machine learning approach and remote sensing indices to identify the degradation of hydraulic structure and surrounding vegetation. The percentage of degradation on the hydraulic structure facility is then calculated and used as an indicator of the deterioration severity. Then, the geometric correction of the inspection images and the quantification of the degraded areas are performed. The multi-period inspection images are orthorectified according to the reference DSM, and thus the degraded areas can be quantified through 3D model reconstruction. And finally, the results are analyzed and evaluated for quality assessment, which include verifying the resultant exterior orientation, image interpretation, and the deterioration area. In this study, we collected inspection images of the hydraulic structure facility around the Yili Bridge. After geometric correction, the positioning accuracy of the image can reach centimeter-level. As for the image interpretation section, we utilized machine learning approach to identify and classify the hydraulic structure facility and remote sensing index to classify vegetation area. The results show that the classification accuracy is able to achieve over 90%. This study is based on the standardized and easy-to-implement process of UAV imagery technology, allowing us to obtain sufficient spatial information from the images to assist in determining the safety of structures. With this we can develop a highly efficient and accurate method to improve the efficiency hydraulic structure facility inspection, which would be useful in river disaster prevention and remediation.