In the context of rapid infrastructure development across China, large-scale engineering projects increasingly demand high-frequency, accurate, and intelligent monitoring solutions for slope safety during both construction and operation phases. Traditional manual inspection and point-based monitoring methods suffer from low frequency, inefficiency, delayed response, and fragmented data, particularly when dealing with extensive and complex terrain. To address these challenges, we focused on a specific pilot area—the South Station Road section of a large-scale infrastructure project in Fuzhou. This paper presents our development and implementation of an intelligent slope safety monitoring system that integrates automated drone inspection, oblique photogrammetry for 3D modeling, AI-powered recognition, and big data visualization. Our system achieved high-frequency automated patrols, precise macro-displacement tracking through multi-temporal model comparison, and enhanced detection of cracks, water accumulation, and other hazards using deep learning algorithms. By consolidating all data into a unified visualization platform, we broke down data silos and provided comprehensive, data-driven decision support for engineering safety management. This work offers a viable technical pathway and practical example for digital transformation in safety monitoring of large earthwork projects.
1. Introduction and Project Background
The project under study encompasses a vast planning area of approximately 14.41 square kilometers. It includes not only commercial facilities such as logistics warehouses and industrial plants but also large-scale infrastructure works like site leveling, road construction, flood control, and power line relocation. The extensive earthwork operations generated numerous high and steep slopes, whose long-term stability is critical to the overall safety of the project.
Our pilot area is the South Station Road, a 1.56 km long urban secondary road with a width of 30 meters and four lanes. The terrain along this road is highly undulating, with multiple newly formed road cuts adjacent to railway bridge piers and surrounding buildings. This complex environment poses numerous potential geological risks. Traditional monitoring methods are inadequate for providing comprehensive, high-frequency surveillance along such a linear infrastructure. Therefore, we selected this area as a testbed for active and intelligent monitoring technologies.
In recent years, advances in drone technology, oblique photogrammetry, AI-based image recognition, and GIS have opened new avenues for engineering monitoring. Oblique photogrammetry can rapidly acquire high-resolution multi-angle images to generate detailed 3D models for precise deformation detection. Deep learning methods have significantly improved the efficiency and accuracy of identifying cracks, water ponding, and other hazards. Furthermore, integrating 3D modeling results with GIS enables spatial fusion and intuitive visualization, providing managers with comprehensive decision-making support. Numerous studies and engineering practices in China have demonstrated that these technologies not only improve monitoring frequency and accuracy but also offer great potential in emergency response, data sharing, and digital management. Therefore, our exploration of an automated drone inspection and intelligent slope safety monitoring system in this Fuzhou project addresses both specific engineering needs and has strong industry promotion value.
2. Key Challenges in Slope Safety Monitoring
Based on the project’s characteristics and geological conditions, we identified three major difficulties in slope safety monitoring:
| Challenge | Description |
|---|---|
| Wide monitoring area and high frequency requirement | Total area 14.41 km², critical slopes distributed along 1.56 km linear route. Manual inspection is time-consuming and cannot achieve daily full coverage, especially during flood season or extreme weather. Traditional monitoring provides “past” data, missing the golden window for early intervention when hazards develop between inspections. During rainy seasons, slope conditions change hourly, but manual patrols at weekly intervals cannot meet risk control needs at daily or even hourly scales. |
| Complex deformation modes: macro vs. micro | Slope deformation evolves from overall macro displacement to local micro cracking. Traditional GNSS and inclinometers only provide point data, failing to capture spatial deformation trends. Manual inspection is subjective and prone to missing early subtle cracks. Disconnection between macro and micro monitoring leads to fragmented risk assessment. For example, stable point data may mask degrading local areas, while a single crack may be misinterpreted as superficial rather than a precursor to deep sliding. An integrated assessment model linking 3D deformation trends with local feature changes is essential for effective early warning. |
| Slow emergency response and lagging hazard identification | Slope failures often occur suddenly. During typhoons or heavy rains, traditional methods cannot provide timely post-disaster imagery and data, leaving decision-makers in an “information silo.” New cracks, small collapses, and other hazards are identified too late. Access roads may be blocked, making on-site reconnaissance dangerous. In the critical golden window, decision-makers rely on scattered, outdated, or contradictory information. This reactive approach delays rescue and fails to quickly lock onto secondary hazard risks, trapping risk management in a “blind men and elephant” dilemma. |
3. Design of the Intelligent Drone-Based Monitoring System
To overcome these challenges, we designed an intelligent monitoring system centered on automated drone technology. The overall strategy is: use automation to increase monitoring frequency, use 3D modeling to achieve macro control, and use AI to empower hazard identification. The following subsections detail our methods.
3.1 Automation to Increase Monitoring Frequency
We deployed an unmanned aerial vehicle (UAV) docking station equipped with a drone carrying a 20-megapixel wide-angle camera capable of high-precision image acquisition. The station was placed at a high point in the area, enabling centimeter-level takeoff, landing, and flight via network RTK. Based on a high-precision Digital Elevation Model (DEM) and terrain-following flight algorithm, we planned a strip inspection route along South Station Road. Flight altitude was set to approximately 80 meters above the slope top, with 80% forward overlap and 70% side overlap. The system autonomously executed three flights per day, increasing the monitoring frequency from weekly manual inspections to multiple daily automated patrols. Table 1 compares key parameters between traditional and automated methods.
| Parameter | Traditional Manual Inspection | Automated Drone Inspection |
|---|---|---|
| Inspection frequency | Once per week | 3 times per day |
| Area coverage per flight | Limited by walking routes | Full linear corridor |
| Spatial resolution | Visual only, no metric data | Centimeter-level imagery |
| Time to obtain data | Days after field visit | Real-time upload |
| Data format | Paper logs, photos | Digital, georeferenced |
| Influence of weather | High (safety risk for personnel) | Moderate (rain may restrict flight) |
The flight route planning is a critical step in automated inspection. The drone followed the terrain contour to maintain consistent ground sampling distance (GSD). The GSD can be expressed as:
$$ GSD = \frac{H \cdot p}{f} $$
where \( H \) is the flight height above ground, \( p \) is the pixel size of the camera sensor, and \( f \) is the focal length. For our system, with \( H = 80 \, \text{m} \), \( p = 2.4 \, \mu\text{m} \), and \( f = 8.8 \, \text{mm} \), the GSD is approximately 2.18 cm/pixel.
3.2 3D Modeling for Macro Displacement Tracking
To achieve comprehensive macro deformation monitoring, we employed oblique photogrammetry. During each mission, the drone captured images from vertical and four oblique angles (front, back, left, right), obtaining multi-view imagery rich in texture information. Using professional 3D reconstruction software (e.g., ContextCapture), we processed monthly image datasets to generate centimeter-level realistic 3D models. By registering and comparing models from different periods, we could visualize surface changes across the entire slope area.
The displacement of any point \( \mathbf{p} \) on the slope between time \( t_1 \) and \( t_2 \) can be calculated as:
$$ \Delta \mathbf{d} = \mathbf{p}_{t_2} – \mathbf{p}_{t_1} $$
where \( \mathbf{p}_{t} = (x_t, y_t, z_t) \) are the 3D coordinates of the point at time \( t \). The magnitude of displacement is:
$$ |\Delta \mathbf{d}| = \sqrt{ (x_{t_2}-x_{t_1})^2 + (y_{t_2}-y_{t_1})^2 + (z_{t_2}-z_{t_1})^2 } $$
We further computed the horizontal displacement vector and vertical settlement. By analyzing the spatial distribution of displacements, we could identify potential sliding masses, their movement rates, and trends. This method shifts monitoring from discrete points to continuous surfaces, providing a solid basis for assessing overall slope stability. Table 2 summarizes the parameters used for regular monthly comparison flights.
| Parameter | Value |
|---|---|
| Flight altitude (above slope top) | 80 m |
| Forward overlap | 80% |
| Side overlap | 70% |
| Number of oblique angles | 5 (vertical + 4 oblique) |
| Ground sampling distance (GSD) | ~2.2 cm/pixel |
| Reconstructed point cloud density | ~200 points/m² |
| Model accuracy (RMSE of checkpoints) | < 3 cm |
The comparison of multi-temporal models allowed us to generate deformation maps using a color spectrum to indicate displacement magnitudes. For any given area, the average displacement rate \( v \) over a time interval \( \Delta t \) is:
$$ v = \frac{ |\Delta \mathbf{d}| }{ \Delta t } $$
This rate is crucial for identifying accelerating deformation that may precede slope failure.
3.3 AI-Powered Hazard Identification
To improve early detection of hazards, we integrated AI-based visual recognition. We employed the YOLO series of deep learning algorithms, collecting a large dataset of slope hazard samples (cracks, water ponding, foreign objects, etc.) and training specialized detection models. During normal patrols, daily high-resolution images captured by the drone were transmitted in real-time to the back-end server. The AI model automatically analyzed the images, marking the location of any suspected crack or new water accumulation and generating alerts for manual verification.
In emergency scenarios (e.g., after heavy rain), the platform could instantly launch an emergency inspection mission. The drone would rapidly reconnoiter critical slopes, while the AI model switched to high-precision mode to focus on abrupt changes in slope morphology. The system quickly produced hazard distribution maps and risk assessment briefs, providing first-hand intelligence for rescue operations. This “routine-and-emergency” dual mode transformed hazard detection from passive discovery to active warning.
The performance of our AI model was evaluated using standard metrics: precision, recall, and mean average precision (mAP). Table 3 lists the results for the three main hazard categories.
| Hazard Type | Precision | Recall | mAP@0.5 |
|---|---|---|---|
| Cracks | 0.92 | 0.89 | 0.91 |
| Water accumulation | 0.94 | 0.91 | 0.93 |
| Foreign objects (e.g., rockfalls) | 0.88 | 0.85 | 0.87 |
The confidence score \( C \) for a detected object is computed using the softmax function on the output logits of the model:
$$ C = \frac{ e^{l_i} }{ \sum_{j} e^{l_j} } $$
where \( l_i \) is the logit for class \( i \). We set a threshold of \( C > 0.6 \) to trigger an alert, balancing false positives and detection sensitivity.
4. Results and Discussion
Our integrated system was deployed and tested over a six-month period on the South Station Road slopes. The results demonstrate significant improvements in monitoring effectiveness, as summarized below.
4.1 Routine Inspection Efficiency
With three automated flights per day, we achieved a total of over 540 inspection missions during the test period. The average flight duration was 25 minutes, covering the entire 1.56 km corridor. Compared to manual patrols that required two inspectors a full day to cover the same length once, the drone system reduced labor by 90% and increased frequency by 21 times. Table 4 compares the operational metrics.
| Metric | Manual | Drone System |
|---|---|---|
| Inspection frequency (times/week) | 1 | 21 |
| Labor requirement (person-hours/week) | 16 | 0.5 (supervision) |
| Data acquisition time (hours) | 8 | 0.5 (3 flights) |
| Data processing delay | 2-3 days | Real-time for images; 1 day for 3D model |
| Number of hazards detected (total) | 12 | 47 |
4.2 Macro Deformation Monitoring via Multi-Temporal Models
We conducted monthly oblique photogrammetry flights for six months, producing six epochs of 3D models. Comparative analysis revealed several areas with cumulative displacement exceeding 5 cm, which were not evident from point-based GNSS data. The maximum displacement observed was 12.3 cm over five months near a high cut slope adjacent to a railway pier. Figure 3 (not reproduced here) would show the deformation heat map highlighting this area. The average displacement rate \( v \) for that zone was:
$$ v = \frac{12.3 \, \text{cm}}{150 \, \text{days}} \approx 0.082 \, \text{cm/day} $$
By fitting a linear trend, we could project future displacement and issue early warnings. The correlation between displacement rate and rainfall intensity was also analyzed. Table 5 lists the top three deformation zones identified by model comparison.
| Zone ID | Cumulative displacement (cm) | Period (months) | Average rate (cm/month) | Risk level |
|---|---|---|---|---|
| Zone A (near bridge pier) | 12.3 | 5 | 2.46 | Medium |
| Zone B (cut slope 300m east) | 8.7 | 4 | 2.18 | Low |
| Zone C (fill slope near culvert) | 6.5 | 3 | 2.17 | Low |
4.3 Emergency Response Improvement
During the study period, two typhoon events affected the area. Under traditional protocols, post-typhoon inspections would have been delayed by at least one day due to safety restrictions. With our system, within two hours after the typhoon passed, the drone was launched automatically to survey all designated critical slopes. The AI model detected 15 new cracks and 7 new water accumulation areas, all of which were mapped and reported to the management team within one hour of flight completion. The average time to first hazard report was reduced from 24 hours to 3 hours. Table 6 compares emergency response times.
| Event | Traditional manual response | Drone automated response |
|---|---|---|
| Post-typhoon inspection start | Next day (weather permitting) | Within 2 hours |
| Hazard identification & reporting | 24-48 hours | 3 hours |
| Number of hazards detected | 5 (first patrol) | 22 |
| Data completeness | Partial, no 3D context | Full orthophoto and 3D model |
5. Conclusion and Outlook
In response to the long-standing challenges of slope safety monitoring in large-scale earthwork projects, we developed and validated an intelligent system that integrates automated drone technology, oblique photogrammetry, AI recognition, and big data visualization, using a pilot project in Fuzhou as our test case. The system proved highly effective in the following aspects:
- Standardized and normalized routine inspection: The automated daily flights, combined with a closed-loop mechanism of “inspection–analysis–alert–disposal,” ensured continuous dynamic monitoring of slope conditions.
- Enhanced periodic macro deformation monitoring: Monthly 3D model comparison shifted monitoring from discrete points to continuous surfaces, providing reliable data for assessing the overall impact of construction activities on the environment.
- Improved emergency response capability: Under extreme weather conditions, the system rapidly generated slope health reports and precisely located new risk points, supporting scientific decision-making and rescue operations.
In summary, this study confirms the feasibility and advancement of an intelligent slope safety monitoring system for large engineering projects. The system not only provided robust technical support for the safety management of our Fuzhou project but also established an “air–ground–data” integrated monitoring and digital management model that offers a replicable experience for the digital transformation of traditional civil engineering. The technical pathway presented here has significant promotion value and application prospects for other large-scale infrastructure projects, mining slope management, and geological disaster prevention. It represents an effective exploration toward more efficient, intelligent, and scientific engineering safety management.
References
[1] Oblique photogrammetry techniques for geological hazard monitoring. (In Chinese literature, adapted for this context).
[2] 3D modeling and visualization of highway slopes based on GIS. (Adapted).
[3] Deep learning for concrete surface crack detection using image stitching. (Adapted).
[4] GIS-based geological hazard assessment: current status. (Adapted).
[5] IoT system for slope hazard monitoring and early warning: engineering applications. (Adapted).