Related Experiment Video
Updated: Apr 22, 2026
08:56
Automatic Image Processing to Determine the Community Size Structure of Riverine Macroinvertebrates
Published on: January 13, 2023
2.5K
Improving performance for multi-category anthropogenic debris detection in river environments by using a
Xiaohan Xu1,2, Cheng Zhang3,4, Hong Huang5
1Department of Civil Engineering, Xi'an Jiaotong-Liverpool University, Suzhou, 215123, China.
Scientific Reports
|April 20, 2026
Summary
Data augmentation using MixUp improves deep learning models for riverine debris detection, with optimal performance varying by model and debris size. YOLOv12n excels with medium augmentation for large objects, while YOLOv11n benefits from low augmentation for small objects.
Area of Science:
- Environmental Science
- Computer Vision
- Machine Learning
Background:
- Accurate detection of riverine debris is crucial for waterway management.
- Deep learning object detection faces challenges in riverine environments due to limited datasets and object variability.
Purpose of the Study:
- To investigate the effectiveness of data augmentation, specifically MixUp, for enhancing deep learning models in detecting riverine debris.
- To develop and utilize a new multi-category dataset with size-tailored annotations for floating debris.
Main Methods:
- A novel multi-category dataset of floating riverine debris was created with scale-specific annotations.
- MixUp data augmentation was applied to YOLOv10n, YOLOv11n, and YOLOv12n models at varying intensities (λ=0.0 to 1.0).
- Model performance was evaluated using mean Average Precision (mAP), focusing on different object sizes and categories.
Main Results:
- YOLOv12n achieved the highest overall performance (71.2% mAP) with medium MixUp intensities (λ=0.3-0.7), improving large-object detection.
- YOLOv11n excelled in small-object detection (e.g., Pomacea canaliculata at 86.1% mAP with λ=0.1) at low-to-medium augmentation levels.
- YOLOv10n performed best without augmentation, indicating incompatibility with MixUp; plastic detection remained high (>95% mAP), while water hyacinth detection was poor (<35% mAP).
Conclusions:
- MixUp augmentation offers model-specific benefits for riverine debris detection, enhancing generalization and performance.
- The choice of augmentation intensity and model architecture significantly impacts detection accuracy for different object sizes and types.
- Further research is needed to address challenges in detecting specific debris like water hyacinth.