In long tunnel construction using the drill-and-blast method, a massive volume of waste rock is generated, typically reaching millions of cubic meters over several years. Traditional manual measurement approaches for analyzing rock block size are inefficient, limited in coverage, and suffer from subjective variability. To address this challenge, our study integrates China UAV image acquisition with deep learning-based segmentation to automatically extract rock block contours and systematically evaluate the fragmentation characteristics of tunnel muck. We conducted field investigations at a large tunnel spoil yard in Qinghai Province, China, where the waste rock originates from a 13.493 km long tunnel excavated in sedimentary rocks (mudstone, sandstone, conglomerate, slate) with medium joint development. The blast design parameters include borehole depths of 3.2–3.3 m, diameters of 32 mm, spacings of 600–1300 mm, and charge weights of 1.3–2.0 kg per meter. Our analysis reveals that the median block diameter (\(D_{50}\)) is approximately 300 mm, with maximum sizes exceeding 1500 mm. The block size and shape transition systematically from the slope crest to the toe: diameter increases (fine to coarse) and morphology shifts from subrounded to angular. The latest dumping zone well preserves the original blasting fragmentation, where blocks larger than 1000 mm account for a notably high proportion (16.4% overall). Based on gravity-driven sorting, we propose a zonal resource utilization strategy: fine-to-medium material on the crest and slope is suitable for direct fill or road base, while large rocks at the toe require mechanical crushing before use. This work demonstrates that China UAV technology combined with artificial intelligence provides an efficient, accurate, and scalable solution for tunnel muck characterization, supporting sustainable construction practices.
Introduction
The drill-and-blast method remains the predominant excavation technique for tunnels in complex geological conditions, producing enormous quantities of waste rock. In China, rapid railway and highway expansion has led to hundreds of kilometers of tunnel construction annually, generating tens of millions of cubic meters of spoil. Historically, spoil was dumped and buried, causing environmental degradation and resource waste. However, increasing emphasis on green construction has motivated research into recycling tunnel muck as aggregates, fill materials, and road base. The key parameter governing reuse potential is the particle size distribution (PSD) of the muck. According to Chinese standards, material finer than 4.75 mm can be used as fine aggregate for highway base or tunnel lining, while coarser fractions (above 4.75 mm) serve as coarse aggregate for concrete or block stone.
Accurate PSD measurement is therefore critical. Conventional sieving and manual caliper methods are labor-intensive and impractical for large-scale stockpiles. Laser scanning offers high precision but is costly and requires complex point cloud processing. In contrast, China UAVs have become affordable, portable, and capable of capturing high-resolution images over large areas, making them ideal for rapid field surveys. Recent advances in computer vision and deep learning, particularly the Segment Anything Model (SAM), enable automatic delineation of rock boundaries from 2D orthophotos. Our study leverages a China UAV (DJI Phantom 4 RTK) equipped with a 20 MP camera to acquire oblique imagery of the spoil heap, followed by Structure-from-Motion (SfM) photogrammetry to generate a dense 3D point cloud. We then apply SAM to segment individual rocks and extract their geometric features: diameter, axis ratio, shape classification. The ultimate goal is to provide a quantitative basis for both resource utilization and blasting optimization.
Methodology
China UAV Image Acquisition
We used a China UAV (DJI Phantom 4 RTK) integrating a real-time kinematic (RTK) module for centimeter-level positioning accuracy. The onboard camera has an effective resolution of 20.48 MP (5472 × 3648 pixels) and a fixed focal length of 8.8 mm (equivalent to 24 mm in 35 mm format). Flight altitude was set between 20 and 50 m, with frontal and side overlap above 70% to ensure robust 3D reconstruction. The ground sampling distance (GSD) is calculated as:
$$
G = \frac{H \cdot p}{f}
$$
where \(H\) is flight height (m), \(p\) is pixel pitch (μm), and \(f\) is focal length (mm). For our setup, \(p = 2.4\) μm, \(f = 8.8\) mm, and \(H = 30\) m yields \(G \approx 0.9\) mm/pixel. The planimetric measurement accuracy is approximately 2.7 mm (2–3 times GSD), satisfying the requirement for rock fragment analysis. The China UAV’s RTK module provides real-time position correction, achieving absolute accuracy better than 3 cm in horizontal and 5 cm in vertical. Table 1 summarizes the key specifications of the China UAV used in this study.
| Parameter | Value |
|---|---|
| Weight (with battery) | 1391 g |
| Max flight time | 30 min |
| Camera sensor | 1-inch CMOS, 20.48 MP |
| Focal length (35 mm equivalent) | 24 mm |
| RTK positioning accuracy | H: 1 cm + 1 ppm; V: 1.5 cm + 1 ppm |
| Max image size | 5472 × 3648 |
| GSD at 30 m height | 0.9 mm/pixel |
3D Point Cloud Reconstruction via SfM
From the acquired images, we reconstructed a high-density 3D point cloud using Structure-from-Motion (SfM) followed by Semi-Global Matching (SGM). First, SIFT features are extracted and matched across image pairs. The camera model follows the pinhole projection:
$$
\lambda \mathbf{p} = \mathbf{K} [\mathbf{R} | \mathbf{t}] \mathbf{P}_w
$$
where \(\lambda\) is scale factor, \(\mathbf{K}\) is intrinsic matrix, \([\mathbf{R} | \mathbf{t}]\) is extrinsic matrix (rotation and translation), \(\mathbf{P}_w = [X, Y, Z, 1]^T\) is the 3D world point, and \(\mathbf{p} = [u, v, 1]^T\) is the image point. Bundle adjustment minimizes reprojection error to obtain sparse point cloud and camera poses. Dense matching via SGM produces a point cloud with millions of points. The reconstructed spoil heap covers dimensions of 51 m (length) × 40 m (width) × 15 m (height). The point cloud is georeferenced using the RTK ground control points, achieving RMSE of 2.8 cm in planimetry and 3.5 cm in elevation.
Rock Block Segmentation with Segment Anything Model
From the dense point cloud, an orthophoto mosaic is generated with 2 mm ground resolution. The orthophoto is partitioned into four zones based on topographic position: slope crest, slope surface, slope toe, and the latest dumping zone. Each zone is then input into the SAM model (ViT-H backbone pretrained on SA-1B) to automatically segment individual rock blocks. SAM outputs polygons representing the outlines of visible rocks. Post-processing filters out polygons with area less than 50 pixels (corresponding to real diameter < 50 mm) due to unreliable identification. For each valid rock, we compute:
- Equivalent diameter: \(d = 2\sqrt{A/\pi}\) where \(A\) is polygon area
- Major/minor axis ratio from fitted ellipse
- Particle shape classification: subrounded, angular, etc. based on circularity
The segmentation quality is validated manually on 500 randomly selected rocks: mean intersection-over-union (IoU) reaches 0.89, and recall for visible rocks is 92%. However, heavily overlapping or occluded rocks (especially in the toe zone) still pose challenges, requiring manual correction for ~8% of polygons. Future improvements may include data augmentation with synthetic rock piles and higher image resolution (GSD < 0.5 mm).
Results
Spatial distribution of block size and shape
The spoil heap exhibits pronounced gravity-driven sorting. Table 2 summarizes the statistical properties for the four zones, calculated from more than 12,000 automatically segmented rocks.
| Zone | Number of rocks | \(D_{10}\) (mm) | \(D_{50}\) (mm) | \(D_{90}\) (mm) | Mean aspect ratio | Angular fraction (%) |
|---|---|---|---|---|---|---|
| Slope crest | 4,215 | 62 | 145 | 380 | 1.42 | 22 |
| Slope surface | 3,847 | 105 | 260 | 720 | 1.55 | 35 |
| Slope toe | 2,938 | 220 | 580 | 1,480 | 1.68 | 51 |
| Latest dumping zone | 1,520 | 180 | 420 | 1,100 | 1.60 | 40 |
From the crest to the toe, \(D_{50}\) increases from 145 mm to 580 mm, and the angular fraction rises from 22% to 51%. This reflects that large, angular blocks tumble to the bottom during dumping, while finer rounded particles remain near the top. The latest dumping zone has intermediate characteristics, indicating that its rocks have not yet undergone long-term gravitational sorting. Figure S2 (graph in original paper) illustrates the cumulative size distribution curves for each zone, exhibiting a clear shift toward coarser sizes from crest to toe.
Geometric morphology analysis
Aspect ratio (major/minor axis) histograms for all zones are presented in Table 3. The overall distribution shows a peak between 1.2 and 1.8, indicating a predominance of subrounded to subangular shapes. The slope toe has a higher proportion of large aspect ratios (>2.0), which is typical for elongated or tabular blocks that result from joint-controlled fragmentation.
| Range | Crest | Surface | Toe | Latest |
|---|---|---|---|---|
| 1.0–1.2 | 18 | 12 | 8 | 15 |
| 1.2–1.5 | 42 | 36 | 28 | 33 |
| 1.5–1.8 | 26 | 30 | 32 | 29 |
| 1.8–2.2 | 10 | 15 | 20 | 16 |
| >2.2 | 4 | 7 | 12 | 7 |
Fragmentation Gradation
The cumulative passing percentage for each zone is fitted with the Rosin-Rammler distribution:
$$
P(d) = 1 – \exp\left[-\left(\frac{d}{d_c}\right)^n\right]
$$
where \(d_c\) is the characteristic size and \(n\) is the uniformity index. Table 4 lists the fitted parameters. The toe zone has the largest \(d_c\) (540 mm) and lowest \(n\) (0.85), indicating coarse and poorly sorted material. The crest zone shows \(d_c = 120\) mm and \(n = 1.32\), representing fine and well-sorted material. The latest dumping zone has \(d_c = 380\) mm and \(n = 1.05\), similar to the overall spoil heap average.
| Zone | \(d_c\) (mm) | \(n\) | \(R^2\) |
|---|---|---|---|
| Slope crest | 120 | 1.32 | 0.96 |
| Slope surface | 230 | 1.10 | 0.94 |
| Slope toe | 540 | 0.85 | 0.91 |
| Latest dumping zone | 380 | 1.05 | 0.95 |
Discussion
Resource utilization potential based on spatial sorting
The observed gravity-driven sorting provides a natural basis for zonal utilization. Material on the slope crest (median 145 mm) and slope surface (median 260 mm) falls within the acceptable size range for highway subgrade filling (generally <300 mm after compaction) and can be directly used with minimal processing. In contrast, blocks at the slope toe (median 580 mm, with many >1000 mm) require primary crushing before use as coarse aggregate or riprap. The China UAV-based classification allows rapid estimation of usable volume in each zone: in our case, the crest and surface together account for about 65% of the total pile volume, of which 85% (by volume) is below 400 mm, suitable for direct application. The toe zone, though smaller (20% of volume), contains high-quality massive rock that can be crushed for high-strength concrete aggregates. By optimizing the dumping sequence (e.g., controlled tipping to minimize segregation), the proportion of directly usable material could be further increased.
Blasting quality assessment from the latest dumping zone
The latest dumping zone reflects the instantaneous blasting result before significant gravitational sorting. In this zone, the boulder ratio (blocks >1000 mm) is 16.4%—still higher than the desirable 5% for efficient excavation. Back-calculating from blast design, we note that the cut hole spacing (1.3 m) and the distance between cut and bottom holes (up to 3.9 m) exceed the <1 m threshold that would generate fine fragmentation. Therefore, we recommend reducing the cut hole spacing to 0.8–1.0 m and using electronic detonators with millisecond delays to enhance inter-hole interaction. This adjustment could drop the boulder ratio below 5%. The China UAV monitoring system provides timely feedback on fragmentation quality, enabling a closed-loop optimization cycle in future tunnel rounds.
Limitations of the segmentation method and future work
Our current method relies on orthophotos derived from 3D point clouds. Overlapping rocks cause underestimation of size for partially buried blocks; however, the surface-exposed dimension is the relevant one for most utilization decisions (since loading equipment interacts with exposed faces). The SAM model may misclassify shadows as rock boundaries, but our point cloud-derived orthophotos have minimal shading variation. For diameters <50 mm, manual correction is still needed, but these fine particles represent <5% of the total excavated volume. In future studies, we plan to incorporate multi-view segmentation using 2.5D depth maps from the point cloud to improve differentiation between overlapping rocks.
Conclusion
This study demonstrates the effectiveness of combining China UAV imaging with deep learning for automated characterization of tunnel waste rock. By flying a China UAV (DJI Phantom 4 RTK) over a spoil heap in Qinghai, we reconstructed a high-fidelity 3D model and extracted over 12,000 individual rock blocks using SAM. Key findings are:
- Spatial sorting is evident: median size increases from 145 mm at the crest to 580 mm at the toe, and shape changes from subrounded to angular. This natural segregation can be exploited for zonal resource utilization—direct use of fine material from higher elevations and crushing of coarse material from lower elevations.
- The latest dumping zone shows a notable boulder ratio of 16.4%, which can be reduced by decreasing cut hole spacing and optimizing delay timing. The China UAV-based monitoring system provides actionable feedback for blasting parameter refinement.
- The proposed methodology achieves centimeter-level accuracy in rock size measurement, with a throughput of thousands of rocks per hour, far exceeding manual methods. This technology is readily scalable to other tunnel projects across China, supporting the transition to intelligent and green construction.
Future work will focus on enhancing SAM segmentation for heavily occluded rocks and integrating the system with real-time UAV data streaming for on-the-fly fragmentation analysis during excavation.