Unmanned aerial vehicle (UAV) systems underwent significant advancements in recent years, which enabled the capture of high-resolution images and accurate measurements, with the tremendous development in artificial intelligence, especially deep learning techniques, Which allows it to be used in the development of Drainage infrastructures that represent a major challenge to confront the flood risks in urban areas and represent a considerable investment, but they are often not as well classified as they should be. In this study, we present an automatic framework for detecting, localizing, and mapping sewer inlets from image clouds acquired by UAVs based on a YOLO CNN architecture. The framework depends on the high image overlap of unmanned aerial vehicle imaging surveys, which were then processed using Structure-from-Motion (SfM) to generate orthomosaic imagery. The framework uses a YOLOv5 model trained to detect and localize sewer inlets in aerial images with a ground sampling distance (GSD) of 3 cm/pixel. Novel object-detection algorithms, including YOLOv5n, YOLOv5s, and YOLOv5x, were compared in terms of the classification and localization of sewer inlets. The approach is evaluated by cross-validating results from an image cloud of 250 UAV images captured over a 0.57 km² study area with 228 sewer inlets. Images with models' performances from the literature, the new YOLO model tested on UAV images in this study demonstrates satisfactory performance, improving both precision and recall. The results show that YOLOv5x offers the best precision  (90%) and recall (92%), whereas YOLOv5n achieved less accuracy in precision and recall (78%) and (80%), respectively. Additionally, increasing image size in the training stage is a very important modification in the model. The study approach has a remarkable ability to detect sewer inlets and can be used to develop the inventory of drainage infrastructure in urban areas.
 
 
Special Issue of AEIC 2024 (Civil Engineering  Session)