Automated Bridge Deck Health Evaluation Aligned with National Bridge Inventory Ratings via Unmanned Aerial Vehicle Imaging and Label-Free Sparse Autoencoder-Based Anomaly Mapping

Bridge deck deterioration poses a critical threat to structural safety and public transportation systems, necessitating scalable and objective inspection methods. This study presents a lightweight, unsupervised anomaly detection framework that leverages unmanned aerial vehicle (UAV)-acquired imagery and sparse autoencoders to evaluate bridge deck surface conditions without requiring labeled training data. High-resolution images captured using a UAV were divided into 64 × 64 patches and processed through a sparse autoencoder trained solely on healthy concrete patches to learn a compact representation of normal surface texture. During testing, reconstruction error was computed for each patch, with elevated errors indicating potential anomalies such as cracks, delamination, or staining. These error values were visualized through heatmaps and aggregated across all patches to derive three condition quantification metrics: average reconstruction error, anomalous area percentage, and normalized severity score. A novel classification scheme empirically mapped these metrics to National Bridge Inventory (NBI) deck condition ratings, offering an interpretable, standardized evaluation of bridge decks. To analyze the model’s robustness and threshold sensitivity, experiments were conducted on eight bridges, showing high agreement with NBI deck condition ratings, achieving up to 87.5% rating classification accuracy. Moreover, threshold sensitivity analysis revealed how rating transitions occur across scoring levels, further highlighting the model’s adaptability. Overall, the proposed approach enables efficient, interpretable, and defect-annotation-free bridge condition assessments, aligning with federal standards while significantly reducing the labor requirements, subjectivity, and data annotation burdens of traditional inspections. It represents a promising step toward scalable, automated infrastructure health monitoring using autonomous aerial systems.