Deep Learning Model Reads Earthquake Orthophotos to Flag Damaged Roads and Bridges

When an earthquake strikes, the first hours of the response phase often hinge on a deceptively simple question: which roads and bridges still work? A new study published in AI in Civil Engineering presents an automated answer to that question, using a deep learning model that scans orthophotos — geometrically corrected aerial or satellite images — and classifies whether the roads and bridges they show have been damaged. The model, built around an EfficientNet-b1 backbone, achieved an accuracy of 94.09 percent on a test set of 220 images drawn from the 2011 Great East Japan Earthquake, a performance the authors describe as sufficient for precisely conducting damage assessment after disasters.

The research team, led by Xianfeng Li and Pang-jo Chun with colleagues from several institutions, set out to address a persistent bottleneck in disaster mapping. Remote sensing has long been recognized as a more efficient way to create disaster maps than traditional field surveys, which demand intensive labor, take prolonged processing time, and are difficult to update and maintain. International organizations such as the International Charter “Space and Major Disasters” and the Copernicus Emergency Management Service already rely on remote sensing to assess damage. Yet the judgment and assessment of infrastructure damage from remote sensing imagery still depend largely on manual analysis, and when large-scale datasets are involved, the overall efficiency of the disaster response stage can be significantly constrained.

Deep learning models have increasingly been applied to close that gap, because they can automatically extract, transform, and abstract features from remote sensing data and process large datasets at speed. Previous efforts have used smartphone-mounted vehicles to collect road damage data, integrated smartphone field data with Synthetic Aperture Radar observations, and applied deep learning to LiDAR point clouds for post-earthquake street network assessment. The authors note that these approaches, while valuable, rely heavily on time-consuming and labor-intensive data acquisition processes. Orthophoto data offers a different route: because it is geometrically corrected, it eliminates the distortions caused by perspective, ensures accurate scale and distances, and enables the acquisition of a large amount of data in a short period of time.

The proposed framework is a binary classification system: given a processed orthophoto, it decides whether the roads or bridges depicted contain damage. Its architecture has three main components. A backbone network extracts hierarchical features from the input image, passing data through successive levels to produce four feature maps of increasing abstraction. A feature pyramid network then constructs a multi-scale feature hierarchy, upsampling and combining feature maps from different network levels so that the model captures both global information and local details. Finally, convolutional heads accomplish the classification, using upsampled and concatenated feature maps passed through fully connected layers to produce the final output.

The multi-scale design matters for this particular task. Finer defects such as cracks and potholes are primarily captured by lower-level feature maps, while larger and more severe structural damage is captured by higher-level feature maps. According to the study, this multi-level collaborative mechanism gives the model superior performance in complex road and bridge damage detection scenarios compared to conventional single-stage detection approaches. The researchers tested three different backbone networks within the same framework — ResNet-152, MobileNet-V3, and EfficientNet-b1 — and compared their performance across the same dataset.

The dataset came from a public collection in the Science Data Bank and contains 820 orthophoto images of the 2011 Great East Japan Earthquake in the Tohoku region. It includes 410 undamaged road or bridge samples and 410 damaged samples. Before training, the team performed semantic segmentation on the images by constructing segmentation masks in which white represents roads and bridges, blue represents water areas, and black represents other areas. By assigning different colors to different areas, the processed images help the model detect road and bridge damage. Of the 820 samples, 600 were used for training — 300 undamaged and 300 damaged — and 220 were held out for testing, split evenly between undamaged and damaged cases. The damaged examples display a variety of forms, including obstructions, submersion, partial missing sections, and structural collapse. Because the linear characteristics of bridges and roads make it feasible to assess their damage together, the model classifies both types of infrastructure with a single decision, and the authors deliberately did not attempt to categorize specific damage types given the high complexity and wide variety of damage forms.

Performance was evaluated with confusion matrices and standard metrics. The EfficientNet-b1-based model outperformed the other two backbones across all evaluation measures. Its accuracy of 94.09 percent reflects a low rate of misclassification. Its precision of 98.99 percent indicates that when the model predicts a sample contains damaged roads or bridges, that positive prediction is highly reliable. Its recall of 89.09 percent reflects strong sensitivity in detecting damaged samples, meaning some damaged cases are still missed. The F1 score of 93.77 percent, the harmonic mean of precision and recall, indicates a well-balanced performance that avoids excessive conservatism or oversensitivity. The model also achieved the highest area under the ROC curve among the three backbones, an AUC of 0.99, approaching the maximum value of 1.0, which the authors interpret as high reliability in distinguishing damaged from undamaged roads and bridges with minimal misclassification.

Perhaps the most instructive part of the study is its honest accounting of the errors. The team sorted the misclassified samples of the best model into three categories and examined a representative example of each. In the first category, an undamaged road was incorrectly classified as damaged: the image showed a road section obstructed by trees and spots on the road surface, and the model could not determine whether the obstructed section remained intact or whether the spots were actual damage. The authors attribute this to the lack of social context and identify the limitation of the single data source as the primary reason. In the second category, an image containing two damaged roads and one intact road was classified as undamaged; because the intact road occupied the majority of the image area, the model apparently focused primarily on it and overlooked the damaged roads, a selective attention problem. The third category is the most concerning: an image in which the road was clearly damaged was still classified as undamaged, a critical misclassification the authors say calls for expanding the training dataset and further improving model performance.

These error patterns translate into three stated limitations. First, the model was trained only on orthophoto data and detects damage exclusively from visual information, so when road or bridge segments are occluded, it may classify the obstructions themselves as structural damage. Second, its detection ability is affected by structural size, and it may overly focus on the main region of an image while overlooking small damaged regions. Third, both the model’s performance and the size of the training dataset require further improvement to enhance robustness and generalization. The dataset itself, while drawn from a real catastrophic event, is modest in size by modern deep learning standards, and the binary output means the model can say only whether damage is present, not where it is or what form it takes.

For future work, the authors suggest replacing or extending the binary classification with semantic segmentation or object detection models, which can better identify and locate damaged areas and thereby enhance the capability of damage detection. They also point to the data availability statement accompanying the paper, which lists public repositories for the original data, the results, and the code, allowing other teams to reproduce the benchmarks or test alternative architectures. In the meantime, the study demonstrates the feasibility of integrating orthophoto data acquired from critical transportation infrastructure during disaster events into deep learning models, offering an efficient way to assess the damage condition of roads and bridges during the disaster response stage — the phase in which every minute of delay in mapping a broken network can matter for rescue and evacuation work.

Subject of Research: Automated detection of road and bridge damage in orthophotos using deep learning

Article Title: Automated detection of bridge and road damage in orthophotos using deep learning

Article References: Li, X., Liang, J., Zheng, S., Lin, C., Chen, Y., Li, H., & Chun, P.-J. (2026). Automated detection of bridge and road damage in orthophotos using deep learning. AI in Civil Engineering, 5(1), Article 24. https://doi.org/10.1007/s43503-026-00108-7

Image Credits: AI Generated

DOI: 10.1007/s43503-026-00108-7

Keywords: deep learning, orthophotos, disaster response, bridge damage, road damage, remote sensing, EfficientNet-b1, feature pyramid network, Great East Japan Earthquake, damage assessment, binary classification, computer vision