Acoustic-Based Queen Bee Status Recognition: A Transfer Learning Approach Refinement

Honeybees are indispensable pollinators for agricultural ecosystems, and a colony's stability and reproductive capacity depend critically on the presence of a healthy queen. Acoustic monitoring has emerged as a promising non-invasive, lighting-independent approach for long-term colony observation. However, existing studies have largely been confined to single-apiary datasets or merged datasets from multiple similar apiaries for model training. Moreover, model evaluation has relied primarily on overall performance metrics, with insufficient attention to cross-region generalization and the detection of queen loss, a rare but critical condition. This study systematically investigates three complementary strategies: noise-augmented data diversification, lightweight convolutional neural network (CNN) architecture optimization via comprehensive ablation experiments, and transfer learning with fine-tuning to bridge the domain gap between source and target apiaries. Under cross-apiary evaluation, the proposed approach achieves an accuracy of 92.79%, a negative-class F1-score of 0.7900, and a negative-class recall of 0.7834 when only limited target-domain training samples are available. With full target-domain training data, the same strategy further attains an accuracy of 95.05%, a negative-class F1-score of 0.8596, and a negative-class recall of 0.8733. t-distributed Stochastic Neighbor Embedding (t-SNE) visualization demonstrates that noise augmentation effectively expands sample diversity in the feature space, while Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps confirm the successful transfer of source-domain acoustic features to the target domain. This work provides a practical approach for deploying acoustic-based queen status monitoring across diverse apiaries with minimal local data collection.