Computer Vision-based YOLO11 for automated damage assessment of public building roofs supporting maintenance management
Year: 2025
Type: conference-paper
Abstract: Maintenance Management (MM) of aging buildings is a critical challenge due to complex maintenance operations and limited resources. Digital technologies, such as Unmanned Aerial Systems (UAS), have improved inspection efficiency and accuracy, while Artificial Intelligence (AI) enhances the reliability of automated analysis of UAS data. However, there is a gap in the literature regarding operational methods and systems that effectively integrate these technologies to provide timely, actionable information for decision-making. Moreover, there is a lack of studies integrating multiple damage classes and roof typologies for automated roof inspections across diverse scenarios. Therefore, this study aims to analyze the application of Convolutional Neural Networks (CNN) to detect roof damage in aging public buildings using UAS-captured images. A case study involved collecting 7,937 drone images from 25 aging buildings, with 7,210 containing roof defects. A YOLO11-based model was developed to automate the recognition of ten damage classes. The top five classes with the highest precision were Biofilm accumulation, Extra Tile, Open or damaged trapdoor, Slab Crack, and Roof Debris. The model showed improvement during training, but dataset imbalance affected its performance, causing it to favor classes with more samples or clearer features. This issue resulted in poorer performance for minority classes, increasing false-negative rates and limiting generalization. Future studies may address dataset imbalance through data balancing, class weighting, or reorganizing the dataset into multiple models. Also, involving experts in roof maintenance could further refine the model's ability to prioritize critical defects.
Source: CIB Conferences
Institution: Universidade Federal da Bahia
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4411472649