We conducted a comprehensive study to address the challenge of efficiently characterizing rock block fragmentation in tunnel spoil produced by drill-and-blast construction. Our work, based on the Dangshun Tunnel project in Qinghai Province, China, introduces an innovative methodology that integrates China drone technology with deep learning algorithms for automated rock block identification and geometric analysis. Traditional manual measurement methods are inefficient and limited in coverage, failing to meet the demands of modern large-scale tunnel engineering. To overcome these limitations, we deployed a China drone equipped with a high-resolution camera to capture multi-view imagery of the spoil yard. The acquired images were processed using Structure from Motion (SfM) and Multi-View Stereo (MVS) algorithms to generate high-precision 3D point cloud models and orthophoto maps. Subsequently, we applied the Segment Anything Model (SAM) for automated extraction of individual rock block contours, enabling a detailed analysis of block size distribution, geometric morphology, and spatial variability.
Our findings reveal that the median rock block diameter (D50) is approximately 300 mm, with maximum sizes exceeding 1500 mm. We observed a clear gravity-driven sorting effect within the spoil pile. From the slope crest to the toe, the diameter of rock blocks progressively increases from fine to coarse, while their morphological characteristics transition from sub-rounded to angular. In the most recently dumped areas, the rock blocks largely retain their original post-blast fragmentation state. However, we identified an anomalously high proportion of large blocks with diameters exceeding 1000 mm in these areas. We attribute this to suboptimal blasting parameters, specifically the spacing between cut holes and floor holes, which exceeds 1000 mm. To improve fragmentation and subsequent resource utilization, we recommend optimizing the charging structure and reducing the blast hole spacing. Based on the observed gravity-induced sorting, we propose a zoned resource utilization strategy: fine and medium-grained materials from the slope crest and face can be directly used as fill or road construction materials, while large blocks concentrated at the slope toe should be crushed for recycling.
Engineering Background and Methodology
The Dangshun Tunnel is a critical control project on the Xining-Chengdu high-speed railway in Qinghai Province, China, with a total length of 13.493 km. The surrounding rock mass primarily consists of Neogene mudstone interbedded with sandstone and conglomerate, as well as Triassic sandstone, slate, and conglomerate, classified as Grade III to IV. The tunnel was excavated using the drill-and-blast method, and the resultant spoil was transported to a designated yard near the tunnel portal. The blasting design parameters, crucial for understanding fragmentation, are summarized in Table 1 below.
| Parameter | Value/Range |
|---|---|
| Charge per hole | 1.3 – 2.0 kg/m |
| Blasthole depth | 3.2 – 3.3 m |
| Blasthole diameter | 32 mm |
| Blasthole spacing | 600 – 1300 mm |
| Cut hole / floor hole spacing | > 1300 mm (up to 3900 mm) |
Our methodology for acquiring and processing data integrates a China drone with state-of-the-art photogrammetry and computer vision techniques. We deployed a DJI Phantom 4 RTK China drone, which incorporates a Real-Time Kinematic (RTK) module for centimeter-level positioning accuracy. The drone captured a sequence of high-resolution images (20 MP) from a flight altitude of 20 to 50 meters. The ground sampling distance (GSD) was calculated using the following formula:
$$ G = \frac{Hp}{f} $$
Where H is the flight altitude, p is the physical pixel size, and f is the focal length. For our parameters, the GSD was calculated to be 0.9 mm, corresponding to a planimetric measurement accuracy of approximately 2.7 mm, which is sufficient for detailed block measurement. The acquired images were processed using the Structure from Motion (SfM) algorithm to reconstruct the camera poses and a sparse 3D point cloud. The core pinhole camera model is represented by:
$$ \lambda P = K [R | t] P_w $$
Here, Pw represents a 3D point in world coordinates, P is its projection onto the image plane, K is the camera intrinsic matrix, [R|t] is the extrinsic matrix representing rotation and translation, and λ is a scale factor. Following SfM, a dense point cloud was generated using the Semi-Global Matching (SGM) algorithm. This process successfully reconstructed a large-scale 3D model of the spoil yard, measuring 51 m in length, 40 m in width, and 15 m in height. Finally, an orthophoto map was generated from this dense model.
For the intelligent extraction of individual rock blocks, we applied the Segment Anything Model (SAM) to the orthophoto images. SAM is a deep learning model pre-trained on a massive dataset of images for promptable segmentation tasks. We partitioned the orthophoto into four distinct zones based on spatial characteristics and deposition history: the slope crest zone, the slope face zone, the slope toe zone, and the latest dumping zone. The segmentation process allowed us to automatically extract the polygonal contours of visible rock blocks in each zone. The model’s performance and limitations were carefully evaluated; results showed high accuracy for isolated and clearly visible blocks but some challenges with overlapping or heavily occluded blocks.
Analysis of Spoil Fragmentation Characteristics
Our analysis focused on three key aspects of the spoil material: spatial distribution, geometric morphology, and block size distribution. The results revealed distinct patterns that have significant implications for resource utilization. For the spatial distribution, we quantified the rock block size distribution for each zone. The cumulative distribution curves demonstrated a clear trend: the slope crest was dominated by fine and medium-sized particles, with over 80% of blocks being smaller than 100 mm. In contrast, the slope toe zone had a significantly coarser distribution, with less than 90% passing the 1000 mm sieve. This pattern is a direct result of gravity-driven segregation, where larger, heavier blocks are more likely to roll and accumulate at the base of the pile, while smaller particles remain near the top.
We further characterized the geometric morphology of the rock blocks. The major and minor axes of each segmented block were measured, and the aspect ratio and particle shape type were analyzed. The key morphological parameters are summarized in Table 2.
| Morphological Parameter | Overall Spoil | Slope Crest | Slope Toe | Latest Dumping |
|---|---|---|---|---|
| Primary Shape | Sub-rounded to Sub-angular | Sub-rounded | Angular | Sub-angular |
| Typical Aspect Ratio (Major/Minor Axis) | 1.2 – 1.8 | 1.2 – 1.5 | 1.5 – 2.0 | 1.3 – 1.7 |
| Roundedness | High | Very High | Low | Moderate |
Overall, the rock blocks were characterized as sub-rounded to sub-angular in shape. However, a clear spatial variation was observed. Blocks at the slope crest were predominantly sub-rounded and more uniform, likely due to attrition during transport and deposition. In contrast, blocks at the slope toe were larger and more angular, indicative of their relatively undisturbed post-blast condition and limited transport distance. Blocks in the latest dumping zone showed an intermediate morphology, preserving much of their original post-blast angularity.
The block size distribution was the most critical characteristic for our study. The overall distribution was broad, with good continuity, indicating a high potential for resource utilization. The D50 for the entire spoil pile was 300 mm, with a Dmax of over 1500 mm. A Rosin-Rammler distribution fitted the data well, which we parameterized as:
$$ R(D) = 100 \cdot 2^{-(D/D_{50})^n} $$
Where R(D) is the cumulative percentage retained on a sieve of size D, D50 is the median size, and n is a uniformity index. The calculated n value for the overall pile was 0.8, indicating a relatively non-uniform distribution. The large block rate, defined as the percentage of blocks with a diameter greater than 1000 mm, was 16.4% for the entire pile. However, this rate was only 2.5% for the latest dumping zone, which was surprising. A more detailed fractal analysis of the number-size relationship was carried out using:
$$ N(D > d) \propto d^{-D_f} $$
Here, N(D > d) is the cumulative number of fragments with a diameter greater than d, and Df is the fractal dimension. The fractal dimension Df for the different zones varied from 2.1 for the crest to 2.8 for the toe, confirming the higher degree of fragmentation in the upper zones.
Discussion and Resource Utilization Strategy
The results of our study have direct implications for both the evaluation of blasting efficiency and the planning of resource utilization. The gravity-driven sorting effect we observed is a fundamental phenomenon in spoil piles. This natural sorting provides a unique opportunity for implementing a cost-effective, zoned utilization strategy. Instead of processing the entire spoil pile uniformly, materials from different zones can be assigned to specific applications based on their inherent characteristics. Our proposed strategy is detailed in Table 3.
| Zone | Characteristic Fragmentation | Proposed Resource Utilization | Pre-Processing Required |
|---|---|---|---|
| Slope Crest | Fine to medium (D50 ~ 50-150 mm), sub-rounded | Direct fill material, road base, subgrade filler | Minimal (e.g., simple screening) |
| Slope Face | Medium to coarse (D50 ~ 150-400 mm), sub-angular | Road base, concrete aggregate (after crushing), rip-rap | Crushing and screening |
| Slope Toe | Coarse to very coarse (D50 > 400 mm), angular | Crushed aggregate for concrete, large rock fill | Heavy mechanical crushing |
| Latest Dumping | Mixed, preserving original blasting state | Blast quality assessment, dynamic parameter optimization | Not recommended for direct use |
The high large block rate observed in the overall pile, particularly the 16.4% rate, indicates room for improvement in the blasting design. Our analysis suggests that the current spacing between cut and floor holes (up to 1.3 meters, and in the case of the cut hole itself, up to 3.9 meters) is a primary cause of these oversized fragments. To mitigate this, we recommend a series of blasting parameter optimizations, which are summarized in Table 4.
| Parameter | Current Value | Recommended Optimization | Expected Effect |
|---|---|---|---|
| Cut hole / floor hole spacing | Up to 1300 mm | Reduce to 800-1000 mm | Reduce large block generation |
| Burden (resistance line) | ~1000 mm (from context) | Optimize based on rock type, reduce to 750-900 mm | Ensure uniform energy distribution |
| Inter-hole delay timing | Standard millisecond delays | Optimize sequence to improve fragmentation, consider electronic detonators | Improve rock breakage and heave |
| Charging structure | Standard column charge | Consider decked or air-deck charges for better energy usage | Reduce over-crushing near borehole and reduce vibrations |
By implementing these proposed optimizations, the large block rate can be significantly reduced, leading to a more uniform fragmentation. This improvement will not only increase the efficiency of subsequent loading, hauling, and crushing processes but will also enhance the overall resource utilization rate of the tunnel muck. The continuous monitoring of the latest dumping zone using our China drone-based methodology will provide the necessary feedback to dynamically fine-tune these blasting parameters.
Conclusion
We successfully demonstrated the efficacy of integrating China drone technology with advanced deep learning models for the intelligent identification and analysis of rock block fragmentation in tunnel spoil. The application of a China drone as a primary data acquisition platform proved to be efficient and cost-effective for large-scale field surveys. Our key findings are as follows:
1) Gravity-Driven Spatial Sorting: The spoil pile exhibits a clear, gravity-driven spatial sorting pattern. Block size increases from the slope crest to the toe, while the morphology transitions from sub-rounded to angular.
2) Resource Utilization Potential: The wide range of block sizes and well-graded nature of the spoil indicate significant resource utilization potential. A zoned strategy, informed by the observed spatial sorting, is proposed to guide the direct use of fine materials and the targeted crushing of coarse materials.
3) Blasting Parameter Optimization: The high proportion of large blocks in the overall pile, particularly associated with the geometry of cut and floor holes, provides a clear target for blasting optimization. Recommendations to reduce hole spacing and adjust delay times are provided to improve fragmentation.
4) Continuous Monitoring and Feedback: The latest dumping zone, which retains the original blasting state, is an ideal subject for assessing blast quality. Regular aerial surveys using a China drone can provide the necessary data for a dynamic feedback loop to continuously refine blasting parameters, ultimately leading to more efficient tunneling operations and higher rates of sustainable resource utilization.
Despite these successes, our study highlights the limitations of current segmentation models for processing overlapping rock blocks. Future work will focus on optimizing the deep learning model architecture and training with more robust datasets incorporating diverse occlusion scenarios. The ultimate goal is to create a fully autonomous, intelligent, and real-time system for the characterization of tunnel spoil, driving the green and sustainable development of tunnel engineering.