CLASS-IMBALANCE-AWARE DEEP SEMANTIC SEGMENTATION FOR VARROA DESTRUCTOR DETECTION UNDER 780-NM NARROW-SPECTRUM ILLUMINATION

Abstract

Varroa destructor remains a major challenge for managed honey bee colonies, and an imaging system that can locate mites directly on bees could support continuous, non-destructive monitoring. The study 'Towards Varroa destructor mite detection using a narrow spectra illumination' established a useful direction by combining controlled illumination with U-Net segmentation and by identifying 780-nm infrared images as a promising condition for mite analysis. The present work builds on that foundation, but changes the research question from whether 780-nm imaging can support segmentation to how the segmentation pipeline can be made more sensitive to the small and highly imbalanced mite class. A reproducible experiment was assembled from the supplied Colab implementation using 647 image-mask pairs, a fixed 80/10/10 split (517 training, 65 validation, and 65 test images), and four semantic-segmentation architectures: U-Net, Attention U-Net, U-Net++, and DeepLabV3+. All models were trained for 30 epochs with Adam at a learning rate of 1e-4, batch size 1, and a 0.5 image scale. The final objective combined weighted CrossEntropy loss with a Varroa-specific Dice loss, using class weights of 0.25, 0.75, and 3.0 for background, bee, and mite, respectively. Checkpoints were selected by validation mite F1-score rather than by overall loss alone. On the held-out test set, U-Net++ achieved the best pixel-level segmentation among the current architectures, with 85.91% precision, 86.62% recall, 86.26% F1-score, and 75.84% IoU. Attention U-Net achieved 85.70% F1 and 74.97% IoU, while the conventional U-Net control reached 84.09% F1 and 72.55% IoU. DeepLabV3+ was much smaller and faster, requiring 3.77 million parameters and 10.25 ms per forward pass, but its mite recall fell to 32.14%. For historical context, the U-Net result reported in the base study corresponds to approximately 90.14% precision, 76.30% recall, 82.65% F1, and 70.42% IoU when calculated from its published confusion counts. The experiments therefore support class-imbalance-aware optimization as a practical way to move the system toward a more balanced sensitivity regime and identify U-Net++ as the strongest pixel-level architecture in the present benchmark. The findings are intentionally presented as an extension of the base study rather than as a controlled replication, because the dataset split, three-class formulation, loss function, checkpoint criterion, and annotation reconstruction pipeline differ.