Tailings ponds are high‑potential energy hazard sources that contain mine waste, and the unimpeded operation of their drainage channels is critical for flood safety during heavy rainfall. Traditional manual inspection is inefficient, incomplete, and often dangerous in steep terrain. To address this gap, we propose an intelligent identification method that fuses UAV drone photogrammetry with the YOLO deep learning model. This approach enables autonomous detection and quantification of drainage channel blockages, providing precise geospatial information for targeted maintenance.
1. Methodology and Data Acquisition
1.1 UAV Drone Photogrammetry
We used a DJI Matrice 4E UAV drone equipped with a 4/3‑inch CMOS camera (focal length 24 mm) to acquire high‑resolution images over two typical tailings ponds in Northeast and North China. The flight altitude was set to 40 m with a ground sampling distance (GSD) of approximately 5.5 mm. Forward and side overlaps were 80% and 70% respectively to ensure robust three‑dimensional reconstruction. A total of 964 images were captured at Tailings Pond A (November 19, 2024) and 1508 images at Tailings Pond B (August 19, 2025). The images were processed using DJI Terra to generate digital orthophoto maps (DOM) of the pond areas.
1.2 YOLO11m‑OBB Model and Dataset Preparation
For automatic detection of drainage channel blockages in the DOM images, we employed the YOLO11m‑OBB (oriented bounding box) model. This model excels at detecting elongated, non‑horizontal objects such as drainage channels with minimal background noise. The dataset consisted of 1032 images, each annotated using roLabelImg with rotated bounding boxes. All labels were converted to TXT format. The dataset was split 7:3 into training and validation sets.
During training, we used the SGD optimizer with an initial learning rate of 0.01, batch size 16, and a total of 200 epochs with early stopping (patience = 100) and weight decay of 0.0005. The final loss values were: box loss = 1.32, cls loss = 0.96, and dfl loss = 2.30. The model achieved a precision of 84.8% and a recall of 90.3% on the validation set, with the confusion matrix showing: true positives (TP) = 140, false positives (FP) = 25, false negatives (FN) = 15, true negatives (TN) = 130. The performance metrics are defined as:
$$
\text{Precision} = \frac{TP}{TP + FP} = \frac{140}{140 + 25} = 0.848
$$
$$
\text{Recall} = \frac{TP}{TP + FN} = \frac{140}{140 + 15} = 0.903
$$
1.3 Grid‑Based Detection Strategy
To apply the model efficiently over the large dam areas, we tiled the DOM images into regular 30 m × 30 m grid cells using ArcGIS fishnet tools. This cell size was chosen to ensure that each grid contained at least one section of drainage channel while keeping the spatial scale small enough for effective feature extraction. Tailings Pond A was partitioned into 332 grid cells, and Tailings Pond B into 662 grid cells. Each cell was sequentially numbered and processed by the YOLO11m‑OBB model.
2. Results
2.1 Identification of Blockages in Tailings Pond A
The detection process for the 332 grid cells of Pond A took 53.59 seconds. The model identified blockages in nine grid cells: Nos. 2, 28, 30, 47, 51, 70, 72, 72 (two blockages), and 73. The total length of blocked sections was 55.95 m. Table 1 summarises the projected coordinates and lengths of each blockage.
| Grid cell No. | X coordinate (m) | Y coordinate (m) | Blockage length (m) |
|---|---|---|---|
| 2 | ***703.757525 | ***8159.08378 | 8.04 |
| 28 | ***164.610468 | ***8256.25423 | 10.06 |
| 30 | ***210.237790 | ***8247.78159 | 3.09 |
| 47 | ***999.940538 | ***8286.41361 | 6.38 |
| 51 | ***124.071910 | ***8263.51573 | 11.12 |
| 70 | ***912.655116 | ***8301.63274 | 5.06 |
| 72 | ***962.651147 | ***8293.28856 | 5.35 |
| 72 | ***979.584514 | ***8290.21939 | 4.98 |
| 73 | ***995.306088 | ***8313.45779 | 1.87 |
2.2 Identification of Blockages in Tailings Pond B
For Pond B, processing the 662 grid cells took 89.14 seconds. The model detected 12 blockages distributed across grid cells 245, 247, 248 (two blockages), 249, 250, 302, 487, 540, 553, and 571 (two blockages). The total blocked length was 114.91 m. Detailed information is presented in Table 2.
| Grid cell No. | X coordinate (m) | Y coordinate (m) | Blockage length (m) |
|---|---|---|---|
| 245 | ***726.446218 | ***9553.16649 | 11.84 |
| 247 | ***785.462941 | ***9545.27973 | 6.06 |
| 248 | ***811.525869 | ***9544.99185 | 4.31 |
| 248 | ***822.284847 | ***9544.57335 | 15.50 |
| 249 | ***859.843325 | ***9544.55622 | 21.98 |
| 250 | ***886.328263 | ***9541.87537 | 8.89 |
| 302 | ***616.392644 | ***9588.32979 | 3.67 |
| 487 | ***489.593657 | ***9838.34403 | 5.46 |
| 540 | ***318.891211 | ***9966.95396 | 9.49 |
| 553 | ***319.136495 | ***9986.40231 | 19.82 |
| 571 | ***456.120784 | ***0020.42529 | 5.06 |
| 571 | ***457.813933 | ***0012.58401 | 2.83 |
3. Discussion
3.1 Efficiency and Practical Benefits
The detection process for Pond A (332 grids) required only 53.59 seconds using the trained YOLO11m‑OBB model, while Pond B (662 grids) took 89.14 seconds—dramatically faster than manual ground inspection. The UAV drone can autonomously survey the entire dam area without exposing personnel to hazardous slopes. By integrating the DOM geoinformation, we obtained precise coordinates and lengths of each blockage, enabling targeted cleanup operations. This method significantly improves the timeliness and coverage of drainage channel maintenance.
3.2 Multi‑purpose Utilization of UAV Drone Photogrammetry
The high‑resolution DOM generated by the UAV drone can be reused for various safety inspections: dam displacement monitoring, crack detection, dry beach length measurement, and ecological assessment. Therefore, a single UAV drone flight mission can serve multiple monitoring tasks, further enhancing the intelligent safety management of tailings ponds.
4. Conclusions
We have developed an intelligent method that combines UAV drone photogrammetry with the YOLO11m‑OBB deep learning model to autonomously detect drainage channel blockages in tailings ponds. Applied to two case study ponds, the model identified 9 blockages (total length 55.95 m) in Pond A and 12 blockages (total length 114.91 m) in Pond B, with processing times under 90 seconds. The approach provides comprehensive blockage information—location and length—directly from the DOM, overcoming the limitations of manual inspection. This work contributes to the automated safety monitoring of tailings ponds and offers a scalable solution for hidden hazard identification.