Thermal bridges are localized regions of a building envelope where thermal resistance is reduced relative to surrounding materials, resulting in increased heat transfer and surface temperature anomalies. Uncrewed aerial system (UAS)–based thermal imaging offers a scalable approach for identifying rooftop thermal bridges across large building stocks; however, manual interpretation of thermal imagery remains time-consuming, subjective, and difficult to apply at the district scale. This study evaluates the use of lightweight deep learning object detection models for automated rooftop thermal bridge detection from UAS-based thermal imagery, with an emphasis on practical deployment efficiency. The analysis uses the Thermal Bridges on Building Rooftops Version 2 data set, which contains 926 images with a total of 6,927 annotations of thermal rooftop bridges collected using UAS platforms and labeled with bounding boxes representing thermal bridge regions. You Only Look Once version 8 (YOLOv8), YOLOv11, and YOLOv26 object detectors, including nano, small, medium, and large variants, are evaluated using thermal-only imagery to reflect realistic operational constraints and reduce input complexity. A unified data processing and training pipeline is developed, including annotation conversion to YOLO format and a spatially disjoint train–test split based on building blocks to ensure generalizable performance evaluation. Model performance is assessed using standard object detection metrics including mean average precision (mAP), precision, recall, F1 score, and localization stability. Experimental results show that among the evaluated models, YOLOv11-s provides the most balanced overall performance, achieving the highest email protected (0.524), the highest recall (0.494), and a tied-best email protected–0.95 (0.283), while maintaining lower model complexity and faster inference than medium and large variants. YOLOv8-m achieved the highest precision (0.687), indicating stronger confidence in predicted detections and fewer false positives. These results demonstrate that lightweight YOLO-based detectors can achieve competitive performance using thermal-only imagery while maintaining computational efficiency suitable for large-scale building energy inspection workflows.
Samsami et al. (Fri,) studied this question.