As a structural health monitoring engineer working on highway bridges across China, I have long been confronted with the limitations of traditional manual inspection methods. In this paper, I present a systematic approach that integrates China drone technology with deep learning algorithms to achieve automated crack detection, segmentation, and quantification. The proposed method has been validated on a real prestressed concrete continuous girder bridge, and the results demonstrate significant improvements in detection rate, measurement accuracy, and operational efficiency compared to conventional practices.
1. Introduction
Bridge structures in China are subjected to heavy traffic loads, temperature gradients, and environmental erosion over decades of service. Surface cracks, as the most common form of deterioration, directly affect the durability and load-bearing capacity of concrete bridges. Traditional manual inspection relies on bridge inspection vehicles or scaffolding, which poses high safety risks, long operation cycles, and limited coverage of hard-to-reach areas such as box girder soffits and cable anchorage zones. The emergence of China drone technology, with its ability to hover precisely at designated altitudes and carry high-resolution cameras, has opened new possibilities for efficient remote inspection. In this work, I combine multi-rotor China drone platforms with an improved convolutional neural network (CNN) to realize automatic crack segmentation and geometric parameter extraction. I also design a hierarchical monitoring system to track crack evolution over time.
2. Background
2.1 Importance of Bridge Crack Detection
Concrete bridges develop various types of cracks during their service life. Transverse cracks often originate from negative moment zones under live load; longitudinal cracks may be related to tendon corrosion or alkali-silica reaction; and map cracking indicates carbonation or freeze-thaw damage. Once a crack penetrates the cover, corrosion of reinforcement accelerates, leading to loss of effective cross-section and potentially catastrophic failure. Timely and accurate measurement of crack location, width, length, and orientation is essential for rating the technical condition of a bridge and planning maintenance interventions. In China, the vast inventory of aging bridges demands more efficient inspection solutions than the traditional manual approach.
2.2 Potential of China Drone Technology in Bridge Monitoring
Multi-rotor China drones can perform vertical take-off and landing, stable hovering, and low-speed cruising. They can be equipped with visible-light cameras, infrared thermography, or LiDAR sensors. Unlike conventional bridge inspection vehicles, a China drone can access confined spaces under the bridge deck and around cable stays. With real-time video transmission and centimeter-level RTK-GNSS positioning, each image can be geo-tagged for multi-temporal alignment. This capability makes dynamic crack monitoring feasible. China drone industry has matured rapidly in recent years, offering cost-effective platforms with flight times of 25–40 minutes and payload capacities suitable for industrial cameras. These advancements directly contribute to the practicality of the method I present here.
3. Detection Method and Technical Advantages
3.1 Limitations of Traditional Crack Detection Methods
Conventional methods include visual inspection by trained engineers, using crack-width gauges, and employing under-bridge inspection vehicles. Visual inspection is subjective and non-quantifiable. Crack-width gauge measurements are accurate but extremely slow for large areas. Under-bridge vehicles require lane closure and skilled operators; a single medium-span bridge can take several days to inspect. Moreover, all these methods produce paper records that are difficult to organize for spatiotemporal analysis and lifecycle management.
3.2 Advantages of China Drone-Based Detection
The China drone inspection platform shifts the workspace from the bridge deck to a three-dimensional airspace. Flight paths can be programmed according to the bridge geometry – for example, longitudinal passes along the girder centerline for deck soffits, and helical ascent around towers for cable-stayed bridges. Quadcopters or hexacopters maintain stable hover even in winds up to 5–10 m/s. Equipped with three-axis gimbals, the camera captures blur-free images. A single China drone sortie covers several thousand square meters in 25–40 minutes. The collected high-resolution orthoimages can be stitched into a digital surface model. With pixel-scale calibration, non-contact measurement of crack parameters becomes possible. All data are stored digitally, enabling the creation of a bridge disease database for full-life-cycle management.
4. Crack Identification Method Based on China Drone Imagery
4.1 Aerial Data Acquisition
I plan each China drone mission by considering bridge type, required resolution, and airspace regulations. For a typical girder bridge, I adopt a “longitudinal back-and-forth” pattern covering the soffit and web zones at a height of 8–15 m. For cable-stayed or suspension bridges, I add spiral routes around the towers and cables. Camera settings follow the principle of minimal motion blur: shutter speed ≥ 1/800 s, ISO adjusted adaptively. The forward overlap is set to ≥70% and side overlap ≥60% to enable subsequent stitching and stereo matching. A real-time kinematic (RTK) module records the position and attitude of the camera at each exposure moment, providing high-precision exterior orientation elements for geo-referencing.
4.2 Image Processing Techniques for Crack Identification
Raw aerial images undergo preprocessing to correct illumination variation and lens distortion. I apply contrast-limited adaptive histogram equalization (CLAHE) to enhance local contrast and highlight crack edges. Radial distortion correction is performed using a standard model. The enhanced images are fed into a semantic segmentation network for pixel-level crack extraction. To address the problem of thin cracks breaking into segments, I use morphological closing to connect adjacent pixels, followed by skeletonization to extract the crack centerline. Width measurement is conducted via gray-level profiling perpendicular to the centerline, converting pixel dimensions to physical dimensions using the ground sampling distance (GSD).
4.3 Role of Machine Learning in Crack Identification
Deep convolutional neural networks (CNNs) excel in semantic segmentation tasks. I adopt an encoder-decoder architecture. The encoder consists of residual blocks that extract multi-scale feature maps; the decoder uses transposed convolutions for up-sampling and skip connections that fuse features from corresponding encoder stages to recover fine edge details. The loss function combines binary cross-entropy and Dice coefficient:
$$ L = -\frac{1}{N}\sum_{i=1}^{N}\left[ y_i \log \hat{y}_i + (1-y_i)\log(1-\hat{y}_i) \right] + 1 – \frac{2\sum_{i=1}^{N} y_i \hat{y}_i}{\sum_{i=1}^{N} y_i + \sum_{i=1}^{N} \hat{y}_i} $$
where \(N\) is the total number of pixels, \(y_i\) is the ground truth (1 for crack, 0 for background), and \(\hat{y}_i\) is the predicted probability. This compound loss balances pixel-wise accuracy and region-level overlap, effectively handling the class imbalance typical of crack segmentation tasks. The model was pre-trained on a self-built dataset of bridge crack images collected by China drone surveys across multiple provinces, achieving a validation intersection-over-union (IoU) of 0.82.
5. Design and Implementation of the Bridge Crack Monitoring System
5.1 System Architecture
The intelligent monitoring system I built adopts a layered architecture consisting of a data acquisition layer, a data processing layer, and an application service layer. The system framework is outlined in Table 1.
| Layer | Components | Function |
|---|---|---|
| Data Acquisition | Hexacopter China drone, 42 MP industrial camera, RTK-GNSS module, 5.8 GHz digital video link | Capture high-resolution images with geo-tags; real-time preview |
| Data Processing | Ground workstation with GPU, CNN model, morphological algorithms | Image preprocessing (distortion correction, CLAHE), semantic segmentation, skeletonization, width calculation |
| Application Service | Web-based dashboard, database, alert system | Display crack distribution maps, evolution charts, severity assessment, push warnings to mobile devices |
5.2 Sensor Selection and Configuration
The airborne imaging sensor is a full-frame CMOS industrial camera with 42.4 effective megapixels (sensor size 35.9 mm × 24.0 mm). Paired with a 35 mm fixed focal length lens, the field of view is approximately 54°. At a flight height of 10 m, the ground sampling distance (GSD) is about 0.85 mm/pixel, enabling detection of cracks with width ≥0.2 mm. The camera is mounted on a three-axis brushless gimbal with stabilization accuracy better than ±0.01° in pitch, roll, and yaw. The RTK module uses dual-frequency multi-constellation receivers (GPS + BeiDou + Galileo) and achieves planar positioning accuracy of 1 cm + 1 ppm and elevation accuracy of 1.5 cm + 1 ppm.
5.3 Data Collection and Processing Workflow
Field operations follow a “global first, then local” strategy. I first fly at a higher altitude to cover the entire bridge panorama, identify suspected concentrated disease areas, and then lower the altitude for detailed re-imaging. Raw images are stored in RAW format to preserve dynamic range. The processing pipeline includes: (1) image quality screening; (2) geo-tagging using RTK data; (3) distortion correction; (4) contrast enhancement; (5) CNN-based semantic segmentation; (6) morphological post-processing and skeletonization; (7) vectorization and attribute extraction. Using a GPU-accelerated batch script, a single 42 MP image is segmented in ~0.8 s. The vectorized cracks are stored in GeoJSON format with fields: crack ID, center coordinates, length, width, orientation, and timestamp. This can be directly imported into a GIS platform for overlay display.
5.4 Optimization of Crack Monitoring Algorithms
To handle varying illumination conditions on site, I introduce data augmentation during training: random brightness, contrast, and hue jitter. During inference, for large-format high-resolution images, I use a sliding window with overlapping merging to eliminate boundary discontinuities. To improve width measurement accuracy, which is sensitive to skeleton localization errors, I adopt a weighted gray centroid method for skeleton refinement. The average width is calculated as:
$$ W = \frac{1}{M} \sum_{j=1}^{M} d_j \cdot \sqrt{1 + \left( \frac{dy}{dx} \right)_j^2} $$
where \(W\) is the average crack width in mm, \(M\) is the number of sampling profiles along the skeleton, \(d_j\) is the pixel span of the crack perpendicular to the skeleton in the j-th profile, and \((dy/dx)_j\) is the tangent slope at the j-th skeleton point, which corrects the projection error for inclined cracks. Multiplying \(d_j\) by GSD gives the physical width. The slope correction term effectively compensates for the underestimation caused by the foreshortening of non-horizontal cracks.
6. Experiments and Results
6.1 Drone Flight Experiment Design
I selected a 28-year-old prestressed concrete continuous girder bridge located in a city in China. The bridge has 5 spans with a total length of 186 m and a deck width of 12.5 m. The test was conducted on a clear, light-wind day (ambient wind speed ~3 m/s, stable illumination). A hexacopter China drone equipped with a 42 MP industrial camera flew along the longitudinal direction at a height of 8 m, speed 2 m/s, forward overlap 75%, side overlap 65%. In total, 412 valid images were captured covering the soffit, webs, and deck pavement. Six checkerboard targets were placed on the ground as ground control points (GCPs) and surveyed with network RTK achieving planar accuracy ±1.5 cm and elevation accuracy ±2.0 cm.
6.2 Crack Identification and Monitoring Results
The collected images were processed through the trained CNN model. The morphology post-processing and skeleton extraction yielded geometric parameters for each crack. A total of 47 cracks were identified. Table 2 lists typical crack quantification results.
| Crack ID | Component location | Orientation type | Length (mm) | Max width (mm) | Avg width (mm) |
|---|---|---|---|---|---|
| C-01 | Span 2, soffit | Transverse | 1258 | 0.42 | 0.31 |
| C-02 | Span 2, soffit | Transverse | 986 | 0.38 | 0.27 |
| C-03 | Span 3, web | Diagonal | 1542 | 0.56 | 0.41 |
| C-04 | Span 3, soffit | Transverse | 2134 | 0.68 | 0.49 |
| C-05 | Span 4, web | Longitudinal | 876 | 0.35 | 0.24 |
| C-06 | Span 4, soffit | Transverse | 1687 | 0.51 | 0.38 |
| C-07 | Span 5, web | Diagonal | 1123 | 0.44 | 0.33 |
| C-08 | Deck pavement | Map | 3256 | 0.72 | 0.53 |
From Table 2, transverse cracks account for a high proportion, mainly located in the negative moment zones of the girder soffits. The maximum crack width of 0.72 mm occurs in the deck pavement map crack area. Spans 3 and 4 show relatively concentrated disease.
6.3 Performance Evaluation and Comparative Analysis
To evaluate the performance of the proposed method, I compared the China drone-based intelligent detection results with those from a traditional under-bridge inspection vehicle operated by three certified engineers using crack-width gauges. The comparative results on the same crack samples are shown in Table 3.
| Metric | China drone intelligent detection | Manual (bridge inspection vehicle) | Improvement |
|---|---|---|---|
| Crack detection rate (%) | 94.6 | 87.2 | +8.5 |
| Width measurement error (mm) | ±0.05 | ±0.08 | −37.5 |
| Time per span (min) | 12 | 95 | −87.4 |
| Total inspection period (h) | 3.5 | 28 | −87.5 |
| Data digitization rate (%) | 100 | 35 | +185.7 |
| Missed detection rate (%) | 5.4 | 12.8 | −57.8 |
| Worker high-altitude exposure (h) | 0 | 22 | −100 |
The results in Table 3 indicate that the China drone-based method improves the detection rate by 8.5%, reduces the missed detection rate from 12.8% to 5.4%, and achieves width measurement accuracy of ±0.05 mm, better than the manual ±0.08 mm. Per-span inspection time is reduced from 95 min to 12 min, and the total inspection period for the bridge is compressed from 28 h to 3.5 h — an efficiency gain of about 8 times. All data are stored digitally (100% digitization), eliminating the traceability problems of paper records. Most importantly, the China drone method completely eliminates the safety risk of workers working at height (0 hours of high-altitude exposure vs. 22 hours).
7. Conclusion
In this study, I have developed an intelligent bridge crack monitoring system that integrates China drone aerial photography with deep learning algorithms, enabling automated processing from image acquisition to crack quantification. The real-bridge validation demonstrates excellent applicability on various component surfaces, including soffits, webs, and deck pavement. Both operational efficiency and measurement accuracy are significantly improved compared to traditional manual methods. Owing to the limitations of visible-light imaging, the current system is only applicable to surface cracks and cannot detect internal voids or hidden cracks beneath the concrete cover. In addition, the training dataset mainly consists of concrete girder bridges; the generalization performance on steel bridges or masonry arch bridges still needs further verification. Future work will incorporate infrared thermography to extend the detection capability to subsurface defects, and expand the training dataset with multiple bridge types to enhance model applicability.