Recent advances in unmanned aerial vehicles (UAVs) and computer vision have opened new frontiers for automated structural health monitoring (SHM) of bridges. Nonetheless, most existing vision-based approaches rely solely on static field images or synthetic datasets, limiting their ability to capture the progression of damage and generalize to real-world conditions. To overcome this drawback, this study presents a hybrid SHM framework that integrates UAV-based visual inspection with a computer vision method trained on both laboratory and field imagery data. Sequential crack images are collected from laboratory tests on concrete beams subjected to progressive shear loading, while UAV-captured images from in-service bridges are incorporated to capture real-world visual variability present in operational environments. The computer vision-based method integrates a convolutional neural network (CNN) architecture with a long short-term memory (LSTM) network. The novelty of this study lies in bridging laboratory sequential data with UAV-captured field imagery through a CNN-LSTM model, enabling lab-to-field integration for SHM. To quantify the impact of temporal learning, the CNN-LSTM model is benchmarked against purely spatial baselines, such as CNN-only model and a CNN with a MobileNet, demonstrating the benefits achieved when temporal learning is incorporated. The proposed CNN–LSTM framework consistently outperformed both baseline configurations, achieving superior accuracy and robustness using laboratory and field datasets. This lab-to-field integration improves the SHM framework’s robustness to environmental noise and provides a scalable path for real-world SHM deployments using UAVs, even in scenarios lacking high-frequency temporal data.
Salehi et al. (Sun,) studied this question.