Industrial Facilities Rooftop Detection Using Aerial Imagery

Inspection is an essential phase for identifying potential defects and minimizing risks, especially in critical building components. The primary focus of this work is the monitoring of industrial facility rooftops. Rooftops are the main part of any building or facility, and roof defects pose risks to personnel and equipment, therefore regular inspections enable timely repairs or replacements, preserving the proper functionality and safety of industrial facilities. In the past, these inspections were carried out through manual surveys, which are time-consuming, prone to human error, costly, and often hazardous for workers. The advent of new technologies, such as Unmanned Aerial Systems (UAS), has proven to be an effective option for overcoming these limitations. Unmanned Aerial Vehicles (UAVs), commonly known as drones, provide potential as effective tools for infrastructure inspections. UAVs enable visual inspections of structures while limiting the need for direct human involvement, reducing inspection costs and time, and minimizing danger to workers. UAVs also provide access to unique types of urban imagery that have traditionally been difficult to capture. For instance, rooftops of industrial facilities often contain equipment such as pipelines, solar panels, water or oil tanks, and heating, ventilation, and air conditioning systems (HVAC), and some areas may be difficult to access due to steep slopes or blocking obstacles. In recent years, significant developments in deep learning, particularly in Convolutional Neural Network (CNN) architectures and related software and hardware, have helped in the analysis of large datasets, achieving outstanding results in many complex tasks. The methodology of combining UAVs with deep learning offers an effective solution for monitoring rooftops in industrial facilities, especially for complex structures. In this work, we use the potential of this methodology by proposing four state-of-the-art CNN architectures, which are MASK R-CNN, DeepLabV3+, U-Net, and YOLOv8, to process aerial images captured by a UAV for automated segmentation and detection of flat and sloped roofs in a synthetic facility map.