Source-linked AI summary
Acoustic Honeybee Queen-State Detection Under Unseen Conditions
Mahsa Abdollahi, Nico Coallier, Maxime Fraser Franco, Tiago H. Falk
TL;DR
Queen-status assessment remains largely manual and disruptive, while generalization across unseen hives and apiaries and the value of pretrained audio representations remain unresolved. The paper benchmarks acoustic representations and classifiers under independent evaluation splits. Modulation spectrograms perform strongest, particularly on unseen hives, but performance declines for unseen apiaries.
Problem
Queen-status assessment remains largely manual and disruptive, while evidence on generalization across unseen hives and apiaries and pretrained representations remains limited.
Source: §1 · p. 1Method
The study benchmarks conventional acoustic representations, CNNs, and pretrained audio representations, emphasizing hive-independent evaluation.
Source: §1 · pp. 1–2Results
Takeaways & Limitations
Modulation-based features show potential for queenless detection, while cross-operation generalization remains a key challenge.
Source: §4 · p. 4, §5 · p. 4Performance decreases considerably on unseen apiaries; the reduction suggests substantial distribution shifts between operations.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Honeybee queen loss is a major threat to colony health, yet queen-status assessment remains largely manual and disruptive. Acoustic monitoring offers a non-invasive alternative by enabling continuous analysis of hive sounds. In this paper, we benchmark conventional and learned acoustic representations for automated detection of queen absence, comparing task-specific convolutional neural networks with pretrained audio transformers. Experiments are performed on 5,129 audio recordings from 3,285 hives across 47 apiaries collected between October 2024 and August 2026, evaluated using hive- and apiary-independent splits. Results show that models relying on modulation spectrograms achieve the best performance, reaching an Area Under the Receiver Operating Characteristic (AUROC) of 0.81 and Area Under the Precision-Recall Curve (AUPRC) of 0.37 on unseen hives; performance decreases under unseen-apiary evaluation. Overall, our results highlight the promise of modulation-based audio representations for non-invasive queen-status monitoring and highlight the challenge in cross-apiary model generalization.
1. INTRODUCTION
The paper addresses unresolved hive- and apiary-level generalization and the limited evaluation of pretrained audio representations for queen-status classification. It benchmarks conventional features, CNNs, and pretrained representations, emphasizing hive-independent evaluation.
1. INTRODUCTION
- Queen loss disrupts brood production and pheromone signaling and can lead to colony decline if it goes undetected.Manual inspections remain time-consuming and disruptive, motivating continuous, non-invasive monitoring.Source: §1 · p. 1
- Prior work used MFCCs and spectral descriptors with classical classifiers, while deep neural models and mel-spectrogram inputs have also been explored for queen-status classification.Mel-spectrograms and MFCCs have remained effective under noisy field conditions.Source: §1 · p. 1
- Modulation features capture temporal dynamics beyond conventional spectrograms, and prior work has reported improved cross-hive generalization for other monitoring tasks [9].The cited applications include colony-population prediction and Varroa mite detection.Source: §1 · p. 1
- Queen-status research still lacks clear evidence of generalization across unseen hives, apiaries, and recording environments, and pretrained audio representations remain underexplored for this task.The paper frames both gaps as targets for systematic evaluation.Source: §1 · p. 1
- The study compares conventional acoustic representations, CNN-based models, and pretrained audio representations, emphasizing evaluation across unseen hives.The evaluation is intended to test whether learned representations generalize across colonies rather than capture colony-specific signatures.Source: §1 · pp. 1–2
2. PROPOSED APPROACH
The paper compares acoustic representations and classifiers for inferring queen state from hive recordings, including metadata fusion into learned acoustic embeddings. It examines representation characteristics and evaluates CNNs alongside pretrained audio transformers.
2.1. Acoustic representations
- Four inputs span MFCCs with temporal derivatives, log-mel, stacked multi-resolution mel, and modulation spectrograms that characterize temporal fluctuations within mel bands.The default framed modulation representation preserves the temporal sequence of modulation windows.Source: §2.1 · p. 2
- Across 160 matched pairs, queenless colonies were louder in every mel band, by about 3 dB below 150 Hz and 1 dB at the wingbeat peak.In the illustrated comparison, the queenless colony showed more slow modulation (< 2 Hz) across 200–1000 Hz.Source: §2.1 · p. 2
2.2. Classification
- All acoustic representations use a shared CNN architecture, while AST and MSM-MAE transformers operate on raw waveforms through their own log-mel front ends.Time and geographic metadata are projected into embeddings and concatenated with the acoustic embedding before classification.Source: §2.2 · pp. 2–3
3. EXPERIMENTAL SETUP
The dataset comprises in-the-wild beekeeper recordings with queen-state labels from a commercial hive-management platform. Evaluation uses hive- and apiary-independent splits to test generalization across colonies and operations.
3.1. Dataset
- The dataset contains 5,129 labeled recordings from 3,285 hives across 47 apiaries, collected between October 2024 and August 2026; 10.9% are queenless.Each recording is a 15-second mono clip resampled to 16 kHz, captured during routine inspections with handheld devices and possible environmental noise.Source: §3.1 · p. 3
- Queen-state labels are based on beekeeper-entered hive status; automated labels are retained only when reviewed and uncontested, and invalid or unavailable labels are excluded.The binary labels are queenright and queenless.Source: §3.1 · p. 3
3.2. Testing Setup and Figures-of-Merit
- Hive-independent splits keep each hive in one split, while apiary-independent splits hold out entire operations; evaluation reports AUROC, queenless-focused AUPRC, precision, and class recalls.Class-balanced cross-entropy addresses the 10.9% queenless share, and test evaluation uses the checkpoint with the best validation loss.Source: §3.2 · p. 3, Figure 3 · p. 3
4. RESULTS AND DISCUSSION
Modulation spectrograms outperform conventional features for queenless detection on unseen hives, but performance falls under apiary-independent evaluation. Pretrained encoders show mixed results, and 3D modulation adds no meaningful improvement over 2D.
4. RESULTS AND DISCUSSION
- AUROC 0.81 and AUPRC 0.37 for 2D modulation spectrograms exceed MFCCs’ 0.70 and 0.24 under the hive-independent split.The comparison is for queenless detection in Table 3.Source: §4 · p. 4, Table 3 · p. 4
- The 3D modulation representation provides no meaningful improvement over the simpler 2D representation.Source: §4 · p. 4
- AUROC 0.63 and AUPRC 0.26 for the best modulation model under apiary-independent evaluation mark a considerable decline, though modulation remains strongest across both splits.The passage suggests distribution shifts between apiaries, potentially related to hive environments, queen lineage, microphone placement, colony management, or background acoustics.Source: §4 · p. 4
- MSM-MAE reaches AUROC 0.79 and AUPRC 0.32 on hive-independent splits, while AST has almost no queenless recall at 0.01; MSM-MAE also falls to AUROC 0.57 on operation-independent splits.The passage concludes that generic pretrained representations do not necessarily provide an advantage over task-specific modulation representations.Source: §4 · p. 4
5. CONCLUSIONS
Modulation spectrograms provide the strongest performance, particularly for unseen hives, but performance declines under apiary-independent evaluation. Cross-operation distribution shifts remain a challenge for robust generalization.
5. CONCLUSIONS
- Modulation spectrograms consistently provide the strongest performance, particularly for unseen hives.Source: §5 · p. 4
- Performance declines under the more challenging apiary-independent setting, highlighting substantial cross-operation distribution shifts.Source: §5 · p. 4
- The conclusion identifies more robust generalization as a remaining need for acoustic queen-status monitoring.Source: §5 · p. 4