Acoustic dataset for "Dual‐Signal Buzz Pollination Monitoring: How Flight and Floral Vibrations Complement Each Other to Improve Bee Species Identification"

Research hypothesis We hypothesized that the acoustic signatures of bee floral buzzes and flight sounds encode species-specific patterns that can be automatically distinguished using machine-learning methods. In particular, by combining “static” features (event duration, fundamental frequency, and a selected subset of Mel-frequency cepstral coefficients) with “dynamic” or cinematic features (statistical moments, RMS envelope, spectral band-power in predefined bands, and a rolling‐element fault indicator), a Random Forest classifier can reliably separate floral buzz from flight and further discriminate among five bee species. Data description & collection This dataset comprises 200 individual sound events (100 floral floral buzzes, 100 flight buzzes recordings) from five bee species. For each event we provide: Metadata (buzz_data.csv) id: unique integer identifier type: behavior label (“floral” or “flight” buzzes) Acoustic features duration_s: buzz or flight duration in seconds fundamental_frequency (Hz) Fifteen selected MFCC coefficients (e.g. mel_7, mel_20, …, mel_167) All raw audio files (.wav), per-sample time‐amplitude exports (.txt), and acoustic feature tables are provided. Notable findings Using a Random Forest classifier (200 trees) with 1 000 randomized train/test splits (33 % hold-out each), we observe: Floral buzz classification: 82.2 % ± 6.7 % accuracy Flight buzz classification: 90.9 % ± 5.2 % accuracy Combined classification: 95.0 % ± 3.9 % accuracy Feature-importance consistently ranked fundamental frequency, certain MFCC quantiles (e.g. 75th and 97.5th percentiles of MFCC 5 and 11), and specific envelope band-power bands (310–400 Hz, 110–130 Hz, 225–250 Hz) as most discriminative. Interpretation & reuse Researchers can load the CSV tables into Python, R, or other tools to: Reproduce or benchmark acoustic classification models Explore species-specific acoustic patterns via feature analysis Extend the dataset with new recordings or alternative feature-extraction methods Integrate into automated pollinator monitoring pipelines The accompanying GitHub repository (DOI: …) contains all Python scripts for feature extraction and model training. Users should first run the cinematic-feature extraction script on the raw .txt files, then apply the classification script to reproduce our results. Licensing & citation This dataset is released under CC BY 4.0. Please cite the Mendeley Data DOI and our code DOI (Zenodo) when reusing.研究假设 我们提出如下研究假设：蜜蜂访花振鸣声与飞行振鸣声的声学特征图谱蕴含物种特异性模式，可通过机器学习方法实现自动区分。具体而言，将"静态特征"（事件时长、基频以及选定子集的梅尔频率倒谱系数（Mel-frequency cepstral coefficients，简称MFCC））与"动态"或时序特征（统计矩、均方根包络、预设频段的频谱带功率以及滚动元件故障指示器）相结合后，随机森林（Random Forest）分类器能够可靠地区分访花振鸣声与飞行振鸣声，并进一步对5种蜜蜂进行物种判别。 数据描述与采集 本数据集包含来自5种蜜蜂的200条独立声音事件录音（其中100次访花振鸣声、100次飞行振鸣声）。针对每条声音事件，我们提供以下内容： 元数据（buzz_data.csv）： id：唯一整数标识符；type：行为标签（"floral"或"flight"振鸣声） 声学特征： duration_s：振鸣声或飞行时长（单位：秒）；fundamental_frequency：基频（单位：赫兹，Hz）；15个选定的梅尔频率倒谱系数（例如mel_7、mel_20……mel_167） 本数据集同时提供所有原始音频文件（.wav格式）、单样本时域幅值导出文件（.txt格式）以及声学特征表。 重要发现 我们采用包含200棵决策树的随机森林分类器，结合1000次随机训练集/测试集划分（每次保留33%数据作为测试集），得到如下结果： 访花振鸣声分类准确率：82.2% ± 6.7%；飞行振鸣声分类准确率：90.9% ± 5.2%；联合分类准确率：95.0% ± 3.9% 特征重要性分析始终显示，基频、部分梅尔频率倒谱系数分位数（例如MFCC5和MFCC11的75分位、97.5分位）以及特定包络频段功率（310–400 Hz、110–130 Hz、225–250 Hz）为最具判别力的特征。 解读与复用 研究人员可将CSV格式特征表导入Python、R或其他数据分析工具，以开展以下工作： 复现声学分类模型或对其进行基准测试；通过特征分析探索物种特异性声学模式；通过新增录音或采用其他特征提取方法扩展本数据集；集成至自动化传粉昆虫监测流程中 本数据集附带的GitHub仓库（DOI：……）包含所有用于特征提取与模型训练的Python脚本。用户需先在原始.txt文件上运行时序特征提取脚本，再调用分类脚本以复现本研究结果。 授权与引用 本数据集采用CC BY 4.0协议发布。若您复用本数据集，请引用Mendeley Data的DOI以及本代码的Zenodo DOI。