Improving performance for multi-category anthropogenic debris detection in river environments by using a size-tailored annotation approach and data augmentation

Accurate detection of anthropogenic riverine debris is essential for environmental monitoring and sustainable waterway management. While deep learning-based object detection methods have shown promising results in related domains, their application to riverine environments remains challenging due to the scarcity of multi-category annotated datasets and the high variability in target sizes and appearances. To address these issues, this study explores the potential of data augmentation techniques to enhance model generalization and performance. Specifically, a new multi-category dataset of floating debris is developed, incorporating size-tailored annotations to better represent objects at different scales. Building upon this dataset, the effects of MixUp augmentation on YOLOv10n, YOLOv11n, and YOLOv12n are systematically investigated across six mixing intensities (λ = 0.0 to 1.0). The experimental results reveal distinct, model-specific responses to MixUp augmentation. YOLOv12n achieved the highest overall performance (71.2% mAP) at medium MixUp augmentation intensities (λ = 0.3-0.7), with significant gains in large-object detection. YOLOv11n performed best at low-to-medium λ values, excelling in small-object detection like maturing Pomacea canaliculata (86.1% mAP at λ = 0.1), while YOLOv10n proved incompatible with MixUp augmentation, peaking without augmentation. Medium-sized objects like plastics maintained high accuracy (mAP > 95%) across settings, whereas water hyacinth remained challenging (mAP < 35.0%).