Geological hazards in high mountain canyon regions pose significant challenges for traditional manual inspection methods. The steep terrain, dense vegetation, and inaccessible slopes create high safety risks and low efficiency. To address these issues, we developed an integrated hardware-software system for intelligent geological hazard inspection leveraging drone hangar technology, close-up photogrammetry, deep learning image recognition, and a unified management platform. This paper presents the overall scheme, key technologies, algorithm design, and field application at a hydropower station dam site in southwestern China. The results demonstrate that our system enables automated, intelligent inspection of fractures, landslides, and passive protective nets, significantly improving inspection efficiency and reducing operational risk.
The core of our system is the automated drone hangar, which allows unmanned, remote, and scheduled flight operations. We selected the DJI Dock 2 (Matrice 3D set) as the hardware platform. This lightweight hangar integrates wind, rain, temperature, and water level sensors to ensure flight safety. The Matrice 3D drone is equipped with a telephoto camera and a mechanical shutter wide-angle camera (4/3 CMOS, 20 MP), capable of 1:500 high-precision mapping. Combined with close-up photogrammetry, it can acquire sub-centimeter resolution images of geological hazard targets. The drone hangar enables fully autonomous takeoff, landing, charging, and data transmission, eliminating the need for on-site operators and enabling continuous periodic inspection.
Key Technologies
Drone Hangar Automation
The drone hangar serves as the physical base for persistent unmanned aerial operations. It supports remote mission planning, real-time weather monitoring, and autonomous charging. The hangar communicates with the central server via a private network, ensuring data security. We deployed the hangar at the left bank outlet yard of the dam site, within 3 km of all demonstration points. The system automatically cancels missions under adverse weather conditions (e.g., heavy rain, strong wind) and resumes when conditions improve. This level of automation is critical for maintaining a high-frequency inspection schedule for geological hazards.
Close-up Photogrammetry
Close-up photogrammetry captures ultra-high-resolution images by flying the drone very close to the target surface. This technique overcomes the limitations of traditional oblique photogrammetry in steep terrain. The process involves three steps: (1) building a coarse model from previous oblique surveys, (2) designing a fine flight path that maintains constant relative altitude and adjusts camera angles, and (3) acquiring images with ground sample distance (GSD) as small as 5 mm. The resulting three-dimensional model allows detailed analysis of cracks, deformation, and structural features. The fundamental geometric relationship in close-up photogrammetry is given by the collinearity equation:
$$
x – x_0 = -f \frac{a_1 (X – X_S) + b_1 (Y – Y_S) + c_1 (Z – Z_S)}{a_3 (X – X_S) + b_3 (Y – Y_S) + c_3 (Z – Z_S)}
$$
$$
y – y_0 = -f \frac{a_2 (X – X_S) + b_2 (Y – Y_S) + c_2 (Z – Z_S)}{a_3 (X – X_S) + b_3 (Y – Y_S) + c_3 (Z – Z_S)}
$$
where (x, y) are image coordinates, (x_0, y_0) are principal point offsets, f is focal length, (X, Y, Z) are object coordinates, (X_S, Y_S, Z_S) are camera projection center, and a_i, b_i, c_i are rotation matrix elements.
Deep Learning for Hazard Recognition
We employed three deep learning models tailored to different geological hazard types. For fracture detection, we used the UPerNet-Swin architecture, which combines a Swin Transformer backbone with a unified perceptual parsing head. For landslide and passive protective net detection, we adopted the SOLOv2 instance segmentation model. SOLOv2 uses position-sensitive segmentation and dynamic convolution to produce accurate masks. To overcome the small sample problem, we applied data augmentation, web crawling, and transfer learning. For landslides, we pre-trained on a satellite-based loess landslide dataset; for protective nets, we used ImageNet pre-training. The training parameters are summarized in the following table.
| Category | Model | Training Size (px) | Batch Size | Learning Rate |
|---|---|---|---|---|
| Fracture | UPerNet-Swin | 512 | 4 | 0.000015 |
| Landslide | SOLOv2 | 2000 | 2 | 0.0001 (optimal) |
| Passive net | SOLOv2 | 1024 | 8 | 0.001 (optimal) |
The loss function for SOLOv2 includes a focal loss for classification and a Dice loss for mask segmentation. The total loss is defined as:
$$
\mathcal{L} = \mathcal{L}_{cls} + \mathcal{L}_{mask}
$$
$$
\mathcal{L}_{cls} = -\alpha (1-p_t)^\gamma \log(p_t)
$$
$$
\mathcal{L}_{mask} = 1 – \frac{2\sum y_{true} y_{pred}}{\sum y_{true} + \sum y_{pred}}
$$
where p_t is the model probability for the true class, α and γ are focal loss hyperparameters, and y_true, y_pred are binary mask values.
Integrated Management Platform
The geological hazard management platform is built on a self-developed digital twin framework. It integrates five subsystems: (1) drone hangar management (mission planning, real-time monitoring, command flight), (2) image recognition (fracture, landslide, passive net detection with historical comparison), (3) 3D visualization (overlaying point clouds, models, sensor data, and hazard labels), (4) result management (automatic report generation and threshold-based alerting), and (5) system management (user permissions and alert thresholds). The platform provides a Web-based interface accessible from any browser, enabling remote supervision and decision support.
System Design and Workflow
The overall workflow consists of five steps:
- Base Model Construction: Use oblique photogrammetry to create a 3D digital surface model of the entire dam site (upstream 2.5 km, downstream 1 km) as the digital twin base.
- Fine Route Planning: For each key hazard site (e.g., tunnel inlet deformation, debris flow gully, high slope with protective nets), design a dedicated close-up photogrammetry route based on the base model.
- Automated Inspection: The drone hangar executes scheduled missions autonomously, uploading images to the server after each flight.
- Intelligent Analysis: The server automatically performs 3D reconstruction and runs the trained deep learning models to identify and segment hazards. Multi-temporal comparison is conducted to detect changes.
- Alert and Management: Results are pushed to the platform’s hazard dashboard. Alerts are triggered if identified parameters exceed user-defined thresholds (e.g., crack width increase, landslide area change).
The platform also provides a manual review function, allowing engineers to confirm or reject automatic detections and update the hazard inventory. This semi-automatic approach balances efficiency with reliability.
Application and Results
Site Description
The test site is a hydropower station in a deep canyon in southwestern China. Three demonstration areas were selected:
- Deformation body at tunnel inlet: Steep slope with fault zones, fractured rock, and tension cracks.
- Debris flow gully on left bank: Gully mouth only 0.9 km from the dam, containing large boulders and potential debris flow under heavy rain.
- High natural slope with protective measures: Active rockfall hazards; existing passive nets and concrete barriers.
Base Model and Route Planning
We used a DJI M300 RTK with Zenmuse P1 camera to collect oblique images covering the entire dam area. The reconstructed 3D model had a resolution of about 2 cm. Based on this model, we defined fine flight paths for each hazard site with GSD of approximately 5 mm. The drone hangar was installed at a safe location within 2 km of all sites, and the flight paths were stored in the platform’s mission library.
Automated Inspection Execution
Over a three-month test period, the system performed 12 automated inspection missions. Each mission covered the three demonstration areas in a single flight (approx. 30 min). The drone hangar operated reliably, with a mission success rate (flights completed without abort) of 95%. Two missions were cancelled due to heavy rain exceeding the hangar’s weather limits. The images were automatically uploaded and processed. The average turnaround from flight completion to hazard detection result was under 2 hours, including 3D reconstruction (30 min) and inference (90 min on Tesla V100 GPU).
Intelligent Recognition Accuracy
We evaluated the detection performance using standard metrics. For fracture segmentation (UPerNet-Swin), we measured pixel-level IoU, precision, recall, and F1-score at an IoU threshold of 0.5. For landslide and passive net instance segmentation (SOLOv2), we computed average precision (AP) and AP50 (AP at IoU=0.5). Results are summarized in the following tables.
| Model | IoU | Precision | Recall | F1-score |
|---|---|---|---|---|
| UPerNet-Swin | 0.5 | 0.6843 | 0.7768 | 0.7276 |
| Category | Model | AP | AP50 |
|---|---|---|---|
| Landslide | SOLOv2 | 0.6326 | 0.748 |
| Passive net | SOLOv2 | 0.7047 | 0.936 |
The crack detection achieved F1-score of 0.7276, performing better on continuous, shadowed fractures and struggling with fragmented, soil-filled cracks. Landslide detection reached AP of 0.6326, with clear boundaries easier to segment. Passive net detection achieved high AP50 of 0.936, likely due to the regular shape and contrast of nets against the background. These results confirm the feasibility of using drone hangar-based image collection combined with deep learning for routine geological hazard inspections.
Multi-Temporal Comparison and Alerting
The platform automatically aligns consecutive models and computes geometric differences for each detected hazard. For example, the width of a crack at the tunnel inlet deformation body was measured over three months. The system detected a 23% increase in average width, triggering an alert. An engineer verified the change by viewing the overlaid models and updating the risk level. Similarly, a landslide scar at the debris flow gully expanded by 12% after a rain event; the platform flagged this as a potential precursor. The passive net integrity index (ratio of intact area to total area) decreased from 0.98 to 0.87 due to rock impact, prompting maintenance work order generation. The platform’s threshold management module allows administrators to set different alert levels for different hazard types based on historical data and expert knowledge.
Integration and User Experience
The drone hangar and the management platform are fully integrated through API. Users can review inspection results via the 3D interactive dashboard, filter hazards by type and time, and export inspection reports in PDF. The system also supports manual annotation and correction of automatic detections, which feeds back into the training dataset for model improvement. This closed-loop design ensures continuous enhancement of recognition performance.
Conclusion
We have demonstrated a comprehensive intelligent geological hazard inspection system based on drone hangar technology. The integration of automated drone hangar operations, close-up photogrammetry, deep learning recognition, and digital twin management enables efficient, safe, and cost-effective hazard monitoring for large hydropower projects. The field application at a dam site in southwestern China validated the system’s performance: fracture detection F1-score of 72.8%, landslide AP of 63.3%, and passive net AP50 of 93.6%. The drone hangar reduced personnel exposure in steep terrain and allowed high-frequency (weekly) inspections that were previously impractical.
Future work will extend the system to cover more hazard types, such as rockfall source areas and drainage system blockages. We also plan to deploy multiple drone hangars to cover larger regions and integrate real-time sensor networks for early warning. The continuous accumulation of inspection data will further improve model accuracy through online learning, making the drone hangar-based intelligent inspection paradigm increasingly robust and autonomous.