Title: Optimizing Brood Cell Identification in Layer Trap Nests for Hymenoptera: A Trade-off Between Annotation Workload and Taxonomic Breadth

Abstract:

Tracking cavity-nesting wild bees and wasps is essential for advancing biodiversity research and conservation strategies. Layer trap nests (LTNs) have emerged as a powerful instrument for assessing the abundance and species richness of these insects, providing critical data on their nesting behaviors and ecological requirements. Nevertheless, the manual assessment of LTNs to identify and categorize brood cells is notoriously labor-intensive and time-consuming. To overcome these bottlenecks, this study introduces a deep learning framework designed for the efficient detection and classification of brood cells within LTNs.

LTNs pose unique difficulties, primarily due to the dense arrangement of brood cells, which significantly increases the labeling burden per image. Furthermore, the dataset exhibits a pronounced class imbalance, where common species are represented far more frequently than rare ones. Fully labeling the abundant species is not only time-intensive but also worsens the existing data imbalance. Conversely, relying on partial labeling leads to incomplete datasets, which negatively impacts model performance. To alleviate the labeling burden and reduce the adverse effects of unlabeled data, we propose a novel Constrained False Positive Loss (CFPL) strategy. This approach dynamically masks predictions derived from unlabeled samples, ensuring they do not disrupt the classification loss during the training process. Our experimental findings indicate that this method enhances detection capabilities, effectively balances the trade-off between model accuracy and annotation effort, and helps address class imbalance issues.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC