DamCrack: A drone and smartphone image dataset for 2D and 3D damage assessment in concrete dams - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Data Brief . 2026 Mar 30;66:112737. doi: 10.1016/j.dib.2026.112737 Search in PMC Search in PubMed View in NLM Catalog Add to search DamCrack: A drone and smartphone image dataset for 2D and 3D damage assessment in concrete dams Vahidreza Gharehbaghi Vahidreza Gharehbaghi a Department of Civil, Environmental and Architectural Engineering, The University of Kansas, KS 66045, USA Find articles by Vahidreza Gharehbaghi a , Caroline Bennett Caroline Bennett a Department of Civil, Environmental and Architectural Engineering, The University of Kansas, KS 66045, USA Find articles by Caroline Bennett a , Rémy D Lequesne Rémy D Lequesne a Department of Civil, Environmental and Architectural Engineering, The University of Kansas, KS 66045, USA Find articles by Rémy D Lequesne a , Hang Zhao Hang Zhao a Department of Civil, Environmental and Architectural Engineering, The University of Kansas, KS 66045, USA Find articles by Hang Zhao a , Jian Li Jian Li a Department of Civil, Environmental and Architectural Engineering, The University of Kansas, KS 66045, USA b Department of Electrical Engineering and Computer Science, The University of Kansas, KS 66045, USA Find articles by Jian Li a, b, ⁎ Author information Article notes Copyright and License information a Department of Civil, Environmental and Architectural Engineering, The University of Kansas, KS 66045, USA b Department of Electrical Engineering and Computer Science, The University of Kansas, KS 66045, USA ⁎ Corresponding author. [email protected] Received 2025 Oct 25; Revised 2026 Mar 17; Accepted 2026 Mar 26; Collection date 2026 Jun. © 2026 The Author(s) This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). PMC Copyright notice PMCID: PMC13091177 PMID: 42007080 Abstract We present DamCrack, a novel dataset comprising high-resolution images of a concrete dam in Spokane, Washington, USA, captured using a drone—including autonomous flights with overlapping images—and mobile devices to document complex surface deterioration. The dataset includes two damage types: cracks and spalling, with pixel-wise annotations provided in different formats. The inclusion of overlapping aerial imagery enables future photogrammetric 3D reconstruction, while the current 2D image data supports immediate computer vision tasks such as damage detection and segmentation. Some damage exhibits visual characteristics resembling alkali-silica reaction (ASR), including map-cracking patterns and surface discoloration. Captured under diverse environmental conditions—such as varying lighting, camera distances, and camera properties—the dataset specifically addresses challenging real-world scenarios where multiple damage types co-occur. DamCrack provides: (1) standardized benchmarks for 2D damage detection and segmentation algorithms, (2) high-quality imagery to support future 3D reconstruction models, and (3) annotated examples of complex, co-occurring damage patterns that contribute to advancing structural health monitoring research for critical infrastructure. Keywords: Concrete dam, Cracks, Spalling, Structural health monitoring, Deep learning, Photogrammetry, 3D Model Reconstruction Specifications Table Subject Engineering & Materials science Specific subject area Civil Engineering Concrete Damage Crack Segmentation 3D Model Reconstruction Photogrammetry Type of data Raw Images (JPG) Derived data : object- and pixel-level annotation files in Pascal VOC Format [ 1 ] YOLO text labels (TXT) and segmentation masks (PNG). Data collection Images were collected from a concrete dam using smartphone cameras and a DJI Mavic 3 Enterprise drone. Close-range images were captured at distances of approximately 3–8 ft, while aerial images were obtained from multiple altitudes (approximately 30–100 ft) with 70–80 % overlap to support both damage detection tasks and photogrammetric 3D reconstruction. Data source location Spokane, Washington, United States Data accessibility Repository Name: DamCrack: A Drone and Smartphone Image Dataset for 2D/3D Damage Assessment in Concrete Dams (Zenodo) Data Identification Number (DOI): 10.5281/zenodo.17274706 . Direct URL to the data: https://zenodo.org/records/17274707 Access conditions: Open access; License: CC BY 4.0. Related research article None Open in a new tab 1. Value of the Data • Target user groups: DamCrack is valuable for researchers and practitioners in structural health monitoring, civil infrastructure inspection, computer vision, photogrammetry, and digital twin development. This dataset is particularly useful for engineers, infrastructure owners, and researchers developing automated methods for crack and spalling detection, segmentation, and condition assessment for concrete dams and related infrastructure systems. • Data gap: Publicly available damage datasets mainly focus on pavements, roads, bridges, tunnels, or laboratory-scale concrete specimens, while datasets representing real-world deterioration in operational concrete dams remain limited. DamCrack addresses this gap by providing a large-scale dataset collected from an actual reinforced concrete spillway structure exhibiting complex deterioration patterns, including crack branching, coexisting spalling, and surface discoloration visually resembling alkali–silica reaction (ASR)-related damage. • Benchmarking and validation: With 11,288 images and multiple annotation formats, DamCrack provides a practical benchmark for training, testing, and validating deep learning models for damage classification, object detection, and semantic or instance segmentation. The availability of pixel-level masks, YOLO labels, and Pascal VOC annotations supports reproducible evaluation and comparison across different model architectures using metrics such as Intersection over Union (IoU) (Jaccard Index) and the Dice Coefficient (F1 score). Reuse and generalization: Because the dataset was collected using multiple cameras, viewpoints, and stand-off distances, it offers valuable diversity for studying model robustness and generalization. Although derived from a reinforced concrete spillway, the damage characteristics and annotation formats make the dataset suitable for reuse in transfer learning, domain adaptation, and comparative studies involving other concrete infrastructure systems such as pavements, retaining walls, tunnels, bridges, and hydraulic structures. 2. Background High-quality data are essential for training deep learning models in structural health monitoring [ 2 ]. However, publicly available damage datasets are largely concentrated on pavements, roads, bridges, and other general infrastructure, while datasets representing real-world deterioration in concrete dams remain limited. This gap is significant because dam surfaces often exhibit complex crack patterns, coexisting spalling, and challenging field conditions that are not well represented in conventional benchmark datasets. To address this need, the DamCrack dataset provides annotated images of concrete dam surfaces containing cracks and spalling captured under realistic operational conditions. In addition to supporting 2D computer vision tasks such as classification, object detection, and segmentation, the dataset also includes high-resolution overlapping aerial imagery suitable for photogrammetric 3D reconstruction. 3. Data Description 3.1. Data overview DamCrack is a dataset of images collected from the Upriver Dam, a concrete gravity dam situated on the Spokane River in Washington state, USA. The dam was originally constructed in 1894 and later rebuilt in 1936. It stands approximately 38 feet high, spans 725 feet in length, and has a hydraulic height of 36 feet. Its primary function is hydroelectric power generation, producing over 73 million kWh annually, sufficient to power the municipal water pumping system that supplies approximately 150 million gallons of water daily to >230,000 residents [ 3 ]. The dam surface exhibits multiple forms of deterioration, including cracking and spalling, making it a suitable case study for health monitoring and automated damage detection research. The image data were collected using both smartphone cameras and an unmanned aerial vehicle (UAV). Fig. 1 shows Upriver Dam with sample images captured using smartphones and the UAV. Fig. 1. Open in a new tab Upriver Dam and representative images. DamCrack supports multiple applications in structural health monitoring, including damage classification, object detection, and crack segmentation. Fig. 2 illustrates sample outputs corresponding to these tasks. In addition, the aerial images, captured from different altitudes and camera angles with approximately 70–80 % overlap, provide suitable input for photogrammetric 3D reconstruction. This capability enables the dataset to support not only 2D damage assessment, but also 3D model reconstruction and digital twin applications for concrete dam inspection and monitoring. Fig. 2. Open in a new tab Different tasks supported by the DamCrack dataset. Overall, DamCrack is comprised of 11,288 images, including 1288 high-resolution raw aerial images (5280×3956 pixels) acquired using a drone, 6000 image patches (512×512 pixels) for damage classification generated from drone and smartphone imagery, and 4000 annotated image patches (640×640 pixels) for damage detection and segmentation. The latter images are annotated using both bounding boxes for detection tasks and pixel-wise masks for segmentation tasks. 3.2. Directory structure This section describes the structure of the repository folders. Fig. 3 provides an overview of the dataset structure, which is organized into folders representing different image sources and annotation types. The dataset includes manually captured images taken with smartphones; aerial images acquired by drone in both autopilot and manual modes; classification images grouped into three class folders; and additional folders containing images and annotations for object detection and segmentation in multiple formats. For segmentation tasks, two separate folders are provided: one containing original labels and the other containing refined labels. Fig. 3. Open in a new tab DamCrack folder structure (damage detection and damage segmentation use the same set of images). For image classification, a deep learning model was initially trained on 300 images across three classes. The trained model was then used to classify the remaining image patches, after which the results were manually reviewed. Based on this process, 6000 image patches were selected to support classification tasks. The images are organized into three subfolders: “non-crack”, which contains concrete surfaces without visible damage; “crack”, which contains images with at least one visible crack; and “background”, which contains images of objects other than concrete, such as ground, sky, and steel structural components (see Fig. 2a ). For damage detection and segmentation applications, the raw high-resolution images captured by drone and smartphones were divided into smaller image patches of 640×640 pixels. From >83,000 generated patches, a subset of 4000 patches was randomly selected for annotation of cracks and spalling. These annotations include pixel-level segmentation masks and bounding boxes for object detection tasks (see Fig. 2b and Fig. 2c ). Bounding boxes for detection were generated by converting segmentation polygons into rectangular regions using the minimum and maximum coordinate values of the polygon boundaries. Table 1 compares the DamCrack dataset and several other damage datasets. A related dataset developed by the authors, DamSegment [ 9 ], is also included for comparison to highlight ongoing efforts toward expanding dam-specific damage datasets. Unlike many existing datasets, DamCrack was collected from an operational dam structure exhibiting a range of concrete defects, including map cracks with varying severity and depth, as well as spalling and popouts of different sizes. Many of the observed damage patterns resemble those associated with ASR, an important deterioration mechanism in concrete structures whose detection is critical for structural health monitoring [ 4 ]. Table 1. Comparison of crack and damage datasets. Dataset Images Tasks Structure CFD [ 5 ] 118 Segmentation Road CrackTree200 [ 6 ] 206 Segmentation Road DeepCrack [ 7 ] 537 Segmentation Pavement, road, wall SUT-Crack [ 8 ] 130 Segmentation, classification, detection Pavement DamSegment [ 9 ] 3500 Segmentation, classification, detection Dam DamCrack 11,288 Segmentation, classification, detection, 3D reconstruction Dam Open in a new tab 3.3. Annotation formats To support a variety of deep learning models, the dataset annotations are provided in multiple formats as illustrated in Fig. 4 . Fig. 4. Open in a new tab Annotation formats used in the DamCrack dataset. 1. Pascal VOC format (.xml) This format supports object detection tasks by providing detailed metadata, including class labels and bounding box coordinates (top-left and bottom-right corners). It is compatible with models such as Faster R-CNN [ 10 ] and SSD (Single Shot Detector) [ 11 ]. In some implementations, it can also be extended to include segmentation information. 2. YOLO format (.txt) Widely used in recent YOLO models (version 8 and newer) [ 12 ], this format stores annotations in plain text files, where each line represents a single object instance. Each entry includes the class label and normalized bounding box coordinates (center x, center y, width, and height). For instance segmentation tasks, polygon coordinates describing the object boundaries can also be included. 3. Mask format (.png): This format is used for semantic and instance segmentation, where each pixel in the mask image is assigned a class value or instance identifier. It enables pixel-level annotation and supports the computation of evaluation metrics such as Intersection over Union (IoU), which are commonly used for segmentation performance assessment. Segmentation models such as DeepLab [ 13 ], U-Net [ 14 ], and Fully Convolutional Networks (FCNs) [ 15 ] are commonly trained and evaluated using this annotation format. 3.4. Comparative overview of DamCrack To quantitatively describe and compare dataset complexity, three statistical indices were computed from the binary crack annotations, as summarized in Table 2 . Damage density ( DD ) measures the proportion of damaged pixels relative to the total image area across the dataset and therefore reflects the overall spatial prevalence of damage. The damage ratio for images ( DR ) was calculated as the ratio of damage pixels to total pixels within each image. The distribution of these image-level damage ratios was then characterized using the 25th and 75th percentiles, denoted as D R 25 and D R 75 , respectively. These percentiles provide a robust description of how crack coverage varies across images. Table 2. Comparison of dataset complexity. Dataset DD D R 75 D R 25 DCI DCI ‾ CFD [ 11 ] 0.029 0.031 0.019 0.00034 0.158 CrackTree200 [ 12 ] 0.022 0.025 0.016 0.00019 0.090 DeepCrack [ 13 ] 0.035 0.044 0.014 0.00105 0.476 SUT-Crack [ 14 ] 0.013 0.014 0.008 0.00007 0.035 DamSegment [ 15 ] 0.048 0.057 0.026 0.00148 0.675 DamCrack (This study) 0.049 0.060 0.015 0.00220 1.000 Open in a new tab To summarize dataset complexity using a single interpretable measure, a Dataset Complexity Index (DCI) was introduced. This index combines the overall damage density with the variability of damage distribution among images and is defined as: DCI = DD × ( D R 75 − D R 25 ) (1) where DD denotes the overall damage density of the dataset, and D R 25 and D R 75 represent the 25th and 75th percentiles of the image-level damage ratio distribution. The range D R 75 − D R 25 reflects the dispersion of crack coverage across images and therefore characterizes the heterogeneity of damage patterns within the dataset. Higher D C I values correspond to datasets with greater overall damage coverage and larger variability in crack distribution among images, indicating more challenging segmentation conditions and increased dataset complexity. In Table 2 , D C I ‾ represents the normalized D C I value, obtained by dividing each dataset’s D C I by the maximum DCI among the compared datasets. As shown, the DamCrack dataset achieves the highest D C I ‾ , indicating that it exhibits the greatest overall damage complexity in terms of both damage density and spatial distribution. 4. Experimental Design, Materials and Methods 4.1. Image acquisition context and setup Image collection for this dataset was conducted using a DJI Mavic 3 Enterprise UAV equipped with a 20 MP sensor. Images were captured from multiple altitudes in both manual and autopilot flight modes. The dataset includes high-altitude, mid-range, and close-up views ( Figs. 5 , 1b and 1c ). Fig. 5. Open in a new tab Example acquisition distances used during data collection. For close-range UAV inspections, the drone was flown in manual mode and maintained at an approximate stand-off distance of 8 feet from the dam surface ( Fig. 5a ). This distance balanced safe maneuverability near the structure with the need to capture high-detail imagery of cracks and spalling. For photogrammetric processing, an image overlap of 70–80 % was maintained during both automated and manual flights to support accurate 3D reconstruction and surface modeling. In addition, supplementary images of different sections of the dam were captured using multiple smartphone cameras, often at shorter distances where access permitted ( Fig. 5b ). These images provide additional close-range views that enhance the visibility of fine damage features. 4.2. Measurement instruments This section summarizes the specifications of the imaging sensors used for data collection. The device models and representative acquisition metadata, such as image resolution, focal length, aperture, and exposure parameters, are presented in Table 3 . For example, images captured using the Samsung N981U have a resolution of 3024 × 3024 pixels, an f/2.2 aperture, a 2 mm focal length, and automatically determined exposure settings. Table 3. Summary of camera sensors and corresponding image specifications. Camera Sensor Image Resolution (pixels) Focal Length (mm) Image Format Aperture Flash Used Smartphone Apple iPhone 13 Pro 3024×4032 6 JPG f/1.5 No Apple iPhone 11 Pro 3024×4032 4 JPG f/1.8 No Apple iPhone 8 Plus 3024×4032 4 JPG f/1.8 No Samsung N981U 3024×3024 2 JPG f/2.2 No Apple iPhone 11 Pro 3024×4032 4 JPG f/1.8 No Apple iPhone 15 Pro 4284×5712 7 JPG f/1.5 No Drone DJI Mavic 3 Enterprise 5280×3956 12 JPG f/2.8 No Open in a new tab All released images are provided in standard RGB JPG format. Prior to publication, personally identifying metadata, including GPS coordinates stored in EXIF records, were removed to protect privacy and comply with data-sharing policies. Consequently, the publicly distributed dataset does not include location-identification metadata. 4.3. Annotation strategy Image annotation was performed using Roboflow, which enabled pixel-level labelling of cracks and spalling within each image patch ( Fig. 6 ). The annotation strategy was designed to prioritize visual consistency, reproducibility, and compatibility with deep learning models operating on RGB imagery. Fig. 6. Open in a new tab Image annotation using Roboflow. In the annotation protocol, cracks were defined as linear or curvilinear surface discontinuities with a dominant elongated geometry (high aspect ratio; length ≫ width). These included continuous, segmented, and branching traces. In contrast, spalling was defined as localized concrete cover loss with an area-dominant footprint (low-to-moderate aspect ratio), typically characterized by visible evidence of material removal such as exposed aggregate, rough texture, or a missing surface layer. Damage classes were determined based on two criteria: (i) geometry (line-like vs. patch-like), and (ii) material-loss evidence (surface discontinuity without loss vs. observable detachment or material removal). For mixed or transitional cases (e.g., locally widened cracks adjacent to small areas of cover loss), the elongated fissure component was labeled as crack, while the contiguous material-loss region was labeled as spalling only when visually distinguishable. Images containing indeterminate regions, unclear areas, or ambiguous damage patterns were excluded to reduce label uncertainty. Because damage boundaries on concrete surfaces are not always sharply defined, especially for fine cracks, crack edges were interpreted based on the visually identifiable fissure region rather than a strict image-intensity threshold or gradient-based boundary. Fig. 7 presents representative annotation examples from the DamCrack dataset. Fig. 7. Open in a new tab Annotation examples from the DamCrack dataset. A: ASR affected area; B: Construction joint. Top: raw image. Bottom: annotation. 4.4. Human–AI iterative annotation and validation The labelling process employed an iterative human-AI collaboration strategy to improve both annotation accuracy and efficiency [ 16 ]. Initially a subset of 2000 images was manually labelled by structural engineering graduate students (hereafter referred to as annotators) under the supervision of licensed structural engineers to establish an initial ground truth dataset. This dataset was used to train a YOLOv9 model [ 17 ], which generated preliminary annotations for an additional 500 images. The annotators then reviewed and refined these AI-generated annotations, correcting missed cracks and inaccurate damage boundaries. A second-generation YOLOv9 [ 17 ] model was subsequently trained using the expanded dataset and used to produce preliminary annotations for another 500 images. This AI-assisted iterative annotation process continued across multiple generations by adding more images to the dataset through each iteration, progressively improving annotation quality while reducing manual labelling effort. To evaluate annotation consistency and quantify the benefits of Human–AI collaboration, a controlled annotator study was conducted using a representative subset of dam imagery. Five independent annotators labeled 560 image samples across seven conditions, including fully manual (No-AI) and AI-assisted generations (Gen0–Gen5). For this study, the latest generation of annotations (Gen5) was used as the reference ground truth. Annotation agreement, defined as the consistency between an annotator’s segmentation mask and the reference ground-truth mask (Gen5), as well as annotation quality, was measured using the Dice Similarity Coefficient (DSC), which quantifies the overlap between two segmentation masks A and B : D S C = 2 | A ∩ B | | A | + | B | (2) where ∣ A ∩ B ∣ denotes the number of overlapping pixels between the two masks. The DSC ranges from 0 to 1, where 1 indicates perfect agreement. As shown in Fig. 8 , the mean DSC increases consistently across successive model generations, indicating progressively improved annotation consistency and boundary fidelity under Human–AI collaboration. The highest agreement is observed in Gen4 and Gen5, where the AI-generated pseudo-masks contain fewer errors and require minimal manual correction. As a result, annotators increasingly converge toward consistent crack and spalling delineations, reflected by the highest DSC values in the final generations. These results demonstrate that iterative Human–AI refinement improves pseudo-label quality while stabilizing annotation behavior across annotators. Fig. 8. Open in a new tab Annotation refinement performance across model generations. 4.5. Annotation boundary refinement To further improve the accuracy and consistency of the damage boundaries, an additional post-processing refinement step was applied following the initial annotation stage. The Segment Anything Model (SAM) developed by Meta [ 18 ] was used to refine the annotated polygons and generate more precise boundary delineations. SAM enables automatic object segmentation using simple prompts, such as initial points or approximate boundaries. By leveraging this capability, the refined masks better capture the fine geometric details of cracks and spalling regions. Fig. 9 illustrates an example of an annotated image and its corresponding refined segmentation. Fig. 9. Open in a new tab Annotation refinement example. 4.6. Similarity-based image filtering Collecting images using multiple cameras and drones often results in overlapping images and repeated views. This is especially common in photogrammetric data collection, where approximately 70–80 % image overlap is required for reliable 3D reconstruction. Consequently, when generating image patches, there is a high likelihood of producing identical or highly similar images. Such redundancy can bias deep learning models toward specific visual patterns, potentially leading to overfitting. As a result, models may achieve high performance on the training dataset while exhibiting reduced generalization capability. To reduce redundancy in the dataset used for segmentation and object detection tasks, we implemented a similarity-based filtering workflow to identify and remove duplicate or highly similar images. The proposed framework is illustrated in Fig. 10 . First, a pretrained deep learning model is used to extract features and generate embedding vectors. Pairwise similarity between these vectors is then computed using cosine similarity. Image pairs exceeding a similarity threshold of 90 % are considered redundant, and one image from each pair is removed from the dataset. Fig. 11 presents example similarity comparisons between image pairs. Fig. 10. Open in a new tab Image similarity pruning framework. Fig. 11. Open in a new tab Similarity comparison. 4.7. Photogrammetric 3D reconstruction The dataset contains overlapping drone imagery captured from different altitudes and camera angles, providing suitable coverage for photogrammetric reconstruction. To evaluate its potential for high-fidelity 3D modeling and digital twin development, the images in the 3D model folder were processed using Agisoft Metashape to reconstruct the spillway surface. The resulting model has also been used and validated for pixel-level image localization, change detection, and texture-based damage detection within digital twin applications [ 16 , 19 , 20 ]. Fig. 12 illustrates the reconstructed 3D model of the dam generated using a Structure-from-Motion (SfM) workflow based on the DamCrack dataset. Fig. 12. Open in a new tab Photogrammetric 3D reconstruction of the dam using the DamCrack dataset. Limitations Segmentation and classification of cracks were conducted by graduate students under the supervision of professional structural engineers, providing strong domain expertise during the annotation process. However, variations in judgment and interpretation among annotators may introduce minor inconsistencies in the labels, particularly in regions where crack severity or boundaries are ambiguous. Ethics Statement The human–AI annotation study described in this work involved human participants performing image annotation tasks. The study protocol was reviewed and approved by the Institutional Review Board (IRB) of University of Kansas (IRB No STUDY00152304). All participants provided informed consent prior to participation. The study was conducted in accordance with the Declaration of Helsinki and applicable institutional ethical guidelines. No personally identifiable information was collected or included in the released dataset. CRediT Author Statement Vahidreza Gharehbaghi: Data Collection, Methodology, Conceptualization, Supervision, Writing; Caroline Bennett: Data Collection, Supervision, Review; Rémy D. Lequesne: Data Collection, Supervision, Review; Hang Zhao: Review; Jian Li: Data Collection, Methodology, Conceptualization, Supervision, Review. Declaration of Generative AI in Scientific Writing During the preparation of this work the authors employed ChatGPT from OpenAI to improve the readability and language of the paper. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. Acknowledgements The authors acknowledge the financial support for this project provided by the U.S. Army Engineer Research and Development Center (ERDC). We also thank the City of Spokane for their coordination and support during the data collection. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data Availability zenodo.org DamCrack: A Drone and Smartphone Image Dataset for 2D/3D Damage Assessment in Concrete Dams (Original data) . References 1. Everingham M., Van Gool L., Williams C.K., Winn J., Zisserman A. The Pascal Visual Object Classes (VOC) challenge. Int. J. Comput. Vis. 2010;88:303–338. [ Google Scholar ] 2. Gharehbaghi V.R., Bennett C., Lequesne R.D., Zhao H., Li J., Alajlan M., Ghulam M.A., Alsalman A. A critical review on structural health monitoring: definitions, methods, and perspectives. Arch. Computat. Method. Eng. 2022;29(4):2209–2235. [ Google Scholar ] 3. “Upriver Dam,” Spokane City. [Online]. Available: www.spokanecity.org . Accessed: 2026. 4. Nguyen A., Gharehbaghi V., Le N.T., Sterling L., Chaudhry U.I., Crawford S. ASR crack identification in bridges using deep learning and texture analysis. Structures. 2023;50:494–507. [ Google Scholar ] 5. Cui L., Qi Z., Chen Z., Meng F., Shi Y. Pavement distress detection using random decision forests. Proc. Int. Conf. Data Science (ICDS); Sydney, Australia; Aug. 2015. pp. 95–102. [ Google Scholar ] 6. Zou Q., Cao Y., Li Q., Mao Q., Wang S. CrackTree: automatic crack detection from pavement images. Pattern. Recognit. Lett. 2012;33(3):227–238. [ Google Scholar ] 7. Zou Q., Zhang Z., Li Q., Qi X., Wang Q., Wang S. DeepCrack: learning hierarchical convolutional features for crack detection. IEEE Transac. Image Process. 2018;28(3):1498–1512. doi: 10.1109/TIP.2018.2878966. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Sabouri M., Sepidbar A. SUT-Crack: a comprehensive dataset for pavement crack detection across all methods. Data Brief. 2023;51 doi: 10.1016/j.dib.2023.109642. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Gharehbaghi V.R., Bennett C., Lequesne R.D., Zhao H., Li J. DamSegment: a curriculum-structured image dataset for damage assessment of concrete dams. Data Brief. 2026 [ Google Scholar ] 10. Girshick R. Proc. IEEE Int. Conf. Computer Vision (ICCV) 2015. Fast R-CNN; pp. 1440–1448. [ Google Scholar ] 11. Zhang S., Wen L., Bian X., Lei Z., Li S.Z. Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR) 2018. Single-shot refinement neural network for object detection; pp. 4203–4212. [ Google Scholar ] 12. Jiang P., Ergu D., Liu F., Cai Y., Ma B. A review of YOLO algorithm developments. Procedia Comput. Sci. 2022;199:1066–1073. [ Google Scholar ] 13. Chen L.-C., Papandreou G., Kokkinos I., Murphy K., Yuille A.L. DeepLab: semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE Trans. Pattern. Anal. Mach. Intell. 2017;40(4):834–848. doi: 10.1109/TPAMI.2017.2699184. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Ronneberger O., Fischer P., Brox T. Proc. Int. Conf. Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2015. U-Net: convolutional networks for biomedical image segmentation; pp. 234–241. [ Google Scholar ] 15. Long J., Shelhamer E., Darrell T. Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR) 2015. Fully convolutional networks for semantic segmentation; pp. 3431–3440. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Gharehbaghi V.R., Li J., Zhao H., Bennett C.R., Lequesne R.D. AI-powered digital twinning and monitoring of concrete dams with drones and photogrammetry. Proc. Int. Workshop Structural Health Monitoring (IWSHM); Stanford, CA, USA; 2025. [ Google Scholar ] 17. Wang C.-Y., Yeh I.-H., Liao H.-Y.M. Proc. Eur. Conf. Computer Vision (ECCV) 2024. YOLOv9: learning what you want to learn using programmable gradient information; pp. 1–21. [ Google Scholar ] 18. Kirillov A., Mintun E., Ravi N., Mao H., Rolland C., Gustafson L., Xiao T., Whitehead S., Berg A.C., Lo W.-Y., Dollár P., Girshick R. Proc. IEEE/CVF Int. Conf. Computer Vision (ICCV) 2023. Segment anything; pp. 4015–4026. [ Google Scholar ] 19. Zhao H., Gharehbaghi V.R., Bennett C., Lequesne R.D., Li J. Pixel-level image localization for updating 3D digital twins of dams using frequency convolutional networks. Automat. Construct. 2026;182 [ Google Scholar ] 20. Zhao H., Gharehbaghi V.R., Li J., Bennett C., Lequesne R.D. YOLO-DCD: a YOLO-based framework for dam damage change detection and 3D model updating. Proc. Int. Workshop Structural Health Monitoring (IWSHM); Stanford, CA, USA; 2025. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement zenodo.org DamCrack: A Drone and Smartphone Image Dataset for 2D/3D Damage Assessment in Concrete Dams (Original data) . Articles from Data in Brief are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (3.6 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top