Deep Learning Scans Earthquake Damage, But Review Finds the Evidence Shaky

Deep learning is rapidly becoming a candidate tool for assessing earthquake damage to buildings, but a new systematic review finds the field far from ready for dependable field deployment. The review, published in AI in Civil Engineering, analysed 46 primary studies published between 2015 and 2025 on image-based, deep-learning-driven assessment of damage to local structural elements in post-earthquake buildings. Conducted under the PRISMA framework, it concludes that while the research area has grown quickly, progress remains uneven across task formulation, dataset design, evaluation protocols, evidence of model generalisation, and deployment readiness.

The review team screened 402 studies retrieved from digital libraries and, after excluding 356, retained 46 studies that all passed a quality assessment threshold. Rather than treating all models alike, the authors organised the literature into four deep learning task configurations: damage classification, damage segmentation, damage detection, and estimation of quantitative damage severity. Classification assigns a single label to an image, such as a damage state or failure mechanism; segmentation assigns a class to every pixel to map the spatial extent of damage; detection identifies whether damage is present and roughly where, using bounding boxes; and regression estimates a continuous engineering quantity such as a damage severity index.

Classification and segmentation dominated the reviewed literature, while object detection was comparatively rare and regression-based severity estimation was the least represented and weakest substantiated task. The authors attribute this pattern partly to annotation burden: classification needs only image-level labels, but segmentation requires pixel-level masks, detection needs spatial annotations, and regression demands quantitative values with clear physical meaning. The rarity of regression, they argue, likely stems from the difficulty of constructing defensible continuous target variables rather than any inherent limitation of the modelling approach itself.

Model adoption has been led overwhelmingly by pre-trained convolutional neural networks such as VGGNet, ResNet, EfficientNet, and MobileNet, mostly applied through transfer learning from large source datasets like ImageNet. For segmentation, encoder-decoder architectures such as U-Net and DeepLabv3+ prevailed, and detection studies mainly used one- and two-stage detectors including YOLOv8 and Faster R-CNN. A handful of more recent studies introduced transformer-based models, though the review notes the evidence base remains too limited to judge their suitability for post-earthquake tasks. Importantly, the authors caution that adoption frequency should not be read as proof of technical superiority, since it is influenced by dataset size, annotation burden, computational cost, and community familiarity.

Performance reporting revealed a familiar pitfall. Most classification models achieved high accuracy, with several model families clustering above 0.90, but the review stresses that high accuracy alone does not demonstrate strong generalisability, particularly where class imbalance exists or only overall accuracy is reported. Segmentation showed a wider performance spread, consistent with the greater difficulty of dense prediction, and overlap metrics such as Intersection over Union proved more informative than accuracy for sparse crack masks. In detection, Faster R-CNN-based approaches outperformed YOLOv8 and pre-trained CNN pipelines on the reported average precision scales. Regression evidence was very limited, with the most promising evaluation, using a modified VGGNet-based feature extraction model, reporting a best RMSE of 0.022 and R² values up to 0.993 under the reported setting.

The datasets underpinning these models drew pointed criticism. Field photography and web-scraped imagery dominated data acquisition, while laboratory, remote sensing, and synthetic photography were used far less often. RGB colour images overwhelmingly prevailed as the data type, with only a few studies incorporating multimodal inputs such as building information modelling metadata or mechanical properties from laboratory experiments. Many studies used datasets of 1,000 to 5,000 samples, datasets exceeding 10,000 samples were uncommon, and class imbalance was widespread. Image resolutions concentrated below 512 pixels, and resizing pipelines common in classification work may suppress fine features such as hairline cracks. A notable proportion of studies did not even report usable resolution information.

The review also identified structural blind spots. Generic or mixed building elements were clearly overrepresented, followed by walls, masonry elements, and columns, whereas beams, beam-column joints, ceilings, and facades were underrepresented. The authors attribute this to data and observability constraints rather than modelling barriers: wall cracks offer large, well-exposed visual targets, while joint and ceiling damage is harder to photograph because of obstructions, access limitations, and partly hidden failure mechanisms. More concerning, many studies predicting damage levels or severity states did not anchor their labels to recognised assessment guidelines such as GB/T 24335-2009, FEMA P-58, or ASCE 41-17, leaving class labels visually convenient but structurally ambiguous.

Perhaps the review’s sharpest finding concerns generalisation. Transfer learning, data augmentation, synthetic data generation, and multimodal enhancement were widely used, but the authors’ evidence-strength assessment found most studies fell within the limited to moderate range for proving these strategies actually improved robustness. Many reported only internal performance gains or benchmark comparisons, which the review deems insufficient to establish robust generalisation. The strongest evidence came from a small subset of studies that combined enhancement strategies with genuinely unseen evaluation, including testing across multiple earthquake events, multiple unseen datasets, or new unseen noisy imagery. Only a very limited number of studies used cross-validation to guard against data leakage, which is particularly risky when image patches originate from the same parent image.

Deployment readiness remains the field’s weakest link. A vast majority of reviewed studies specified no practical deployment pathway, and the few that moved beyond offline benchmarking produced a tracking and reporting system, a web-based interface for damage classification, a BIM-informed recovery planning workflow, a mobile application, and a repair cost estimation pipeline. Real-time capability was demonstrated only in isolated cases, including one model achieving 16 milliseconds of inference per frame on a GTX 1080Ti GPU, and no reviewed study demonstrated inference on embedded edge hardware under realistic post-disaster conditions. Explainable artificial intelligence was largely absent, and no study established a validated field-ready workflow spanning autonomous capture, low-latency processing, uncertainty-aware interpretation, and engineering decision support.

The review’s verdict is that the field now possesses many core prerequisites but has not combined them into a consistently validated, deployment-oriented framework. Its recommendations include data-efficient models, stronger multimodal learning combining imagery with thermal data, inspector notes, structural metadata, or vibration response, stricter evaluation on unseen field data, structurally grounded labelling protocols anchored to recognised damage assessment guidelines, and integration with real-time, engineer-facing assessment workflows. Until those pieces mature, image-based deep learning for local post-earthquake damage assessment remains, in the authors’ framing, a set of promising research-level demonstrations rather than a dependable tool for the engineers, governments, insurers, and first responders who need it most.

Subject of Research: Deep learning for image-based assessment of structural element damage in post-earthquake buildings

Article Title: Deep learning for image-based structural element damage assessments in post-earthquake buildings: a systematic review

Article References: Abeysuriya, K. G., Ciupala, M. A., Ghorashi, S. A., & Ilki, A. (2026). Deep learning for image-based structural element damage assessments in post-earthquake buildings: a systematic review. AI in Civil Engineering, 5(1), Article 15. https://doi.org/10.1007/s43503-026-00099-5

Image Credits: AI Generated

DOI: 10.1007/s43503-026-00099-5

Keywords: deep learning, post-earthquake damage assessment, structural engineering, systematic review, damage classification, damage segmentation, transfer learning, convolutional neural networks, multimodal data, model generalisation, deployment readiness, PRISMA