Dataset from Danum Valley Conservation Area, Sabah, Malaysia for BirdNET transfer learning exercise

Dataset summary This dataset contains two folders. The first folder 'Trainingdata' contains short audio clips (.wav) of sounds from four categories: 'gibbon.female', 'hornbill.helmeted', 'hornbill.rhino', and 'noise'. Each folder contains 10 individual clips. The second folder 'Testdata' contains one 2-hour sound file and a corresponding Raven selection containing manual annotations of 'gibbon.female', 'hornbill.helmeted' and 'hornbill.rhino'. Passive acoustic monitoring data collection We collected data using first generation Swift autonomous recording units (ARUs) with a microphone sensitivity of −44 (+/−3) dB re 1 V/Pa. We collected acoustic data from one primary conservation area in Sabah, Malaysia: Danum Valley Conservation Area (using 11 recording units from March to July 2018). Danum Valley covers an area of roughly 440 km², and is characterized by lowland dipterocarp forest. Unlike many tropical forest regions, this area is generally considered 'aseasonal' due to its lack of clearly differentiated wet and dry seasons. The ARUs recorded at a sampling rate of 16 kHz, and all recordings were saved in waveform audio (.wav) format, with files of 2-hr duration. We affixed each recording unit to trees approximately 2-m above the ground and recorded continuously over 24 hours. We set the units on a 750 m grid structure. Acoustic data preparation We randomly chose approximately 500 h of recordings from Danum Valley Conservation Area to use to create a training dataset. We used a band-limited energy detector (BLED) to identify potential sounds of interest in the gibbon frequency range. For the BLED detector, we converted the 2-hr recordings into a spectrogram using a 1,600-point (100 ms) Hamming window (3 dB bandwidth = 13 Hz) with 0% overlap and a 2,048-point DFT, with the "seewave" package (Sueur et al. 2008). We then filtered the spectrogram to focus on the desired frequency range, specifically 0.5–1.6 kHz for Northern grey gibbons. For each unique time window in the recording, we determined the total energy across frequency bins which gave a single value for every 100 ms interval. Utilizing the "quantile" function in base R, we established the threshold to delineate signal from noise. Preliminary tests with varied quantile values revealed that the 15th quantile led to optimized recall for our target signal. We considered cases where all time windows between 6 - 20 seconds exceeded the threshold as 'sound events', and this approach resulted in 1,439 unique sound events. The sound events were then annotated by a single observer (DJC) using a custom-written function in R to visualize the spectrograms into the following categories: helmeted hornbills (Rhinoplax vigil), rhinoceros hornbills (Buceros rhinoceros), female gibbons (Hylobates funereus) and a catch-all “noise” category. Data Usage If you use this dataset, please also cite this paper: Clink, D.J., Kier, I., Ahmad, A.H. and Klinck, H., 2023. A workflow for the automated detection and classification of female gibbon calls from long-term acoustic recordings. Frontiers in Ecology and Evolution, 11, p.1071640. https://www.frontiersin.org/articles/10.3389/fevo.2023.1071640/full

数据集概况
本数据集包含两个文件夹。第一个文件夹「训练数据集（Trainingdata）」内包含四类声音的短音频片段（.wav格式）：雌性长臂猿（gibbon.female）、盔犀鸟（hornbill.helmeted）、双角犀鸟（hornbill.rhino）以及噪声（noise），每个类别对应的子文件夹均包含10条独立音频片段。第二个文件夹「测试数据集（Testdata）」包含一段2小时的音频文件，以及配套的Raven选择标注文件（Raven selection），该文件中包含针对雌性长臂猿、盔犀鸟与双角犀鸟的人工标注信息。

被动声学监测数据采集
本研究采用第一代Swift型自主记录单元（autonomous recording units, ARUs）采集数据，其麦克风灵敏度为-44(±3) dB，参考值为1 V/Pa。数据采集地点位于马来西亚沙巴州的丹浓谷保护区（Danum Valley Conservation Area），于2018年3月至7月间布设11台记录单元开展声学数据采集。丹浓谷保护区面积约440平方千米，以低地龙脑香科森林为典型植被。与多数热带林区不同，该区域无明显干湿季交替，因此通常被视为「无季节分化」区域。记录单元的采样率设为16 kHz，所有录音均以波形音频（.wav）格式存储，单文件时长为2小时。每台记录单元被固定在距地面约2米的树干上，实现24小时全天候连续录音，记录单元按750米的网格布局进行布设。

声学数据预处理
本研究从丹浓谷保护区的录音档案中随机选取约500小时的音频数据用于构建训练数据集。采用带限能量检测器（band-limited energy detector, BLED）识别长臂猿频率范围内的潜在目标声音。针对BLED检测器，借助「seewave」R包（Sueur等，2008），将2小时的录音转换为语谱图：采用1600点（100 ms）汉明窗（3 dB带宽=13 Hz），无重叠区间，使用2048点离散傅里叶变换（discrete Fourier transform, DFT）。随后对语谱图进行滤波处理，聚焦于北灰长臂猿的目标频率范围0.5–1.6 kHz。针对录音中的每个独立时间窗口，计算各频率段的总能量，得到每100 ms间隔的单一能量值。利用R基础包中的「quantile」函数设定阈值，以区分信号与噪声。通过对不同分位数取值的预测试验证，15%分位数可使目标信号的召回率达到最优。将6至20秒区间内所有时间窗口的能量值均超过阈值的情况判定为「声音事件」，该方法共提取得到1439个独立声音事件。随后由单一名观察者（DJC）借助R语言自定义脚本，将语谱图可视化并标注为以下四类：盔犀鸟（Rhinoplax vigil）、双角犀鸟（Buceros rhinoceros）、雌性长臂猿（Hylobates funereus）以及通用的「噪声」类别。

数据使用规范
若使用本数据集，请同时引用以下论文：
Clink, D.J., Kier, I., Ahmad, A.H. 与 Klinck, H., 2023. 长期声学录音中雌性长臂猿叫声的自动检测与分类工作流. 《生态学与进化前沿》(Frontiers in Ecology and Evolution), 第11卷, 第1071640号. https://www.frontiersin.org/articles/10.3389/fevo.2023.1071640/full

搜集汇总

背景与挑战

背景概述

该数据集来自马来西亚丹浓谷自然保护区，用于BirdNET迁移学习练习，包含训练和测试数据：训练数据有四个类别（雌性长臂猿、盔犀鸟、犀鸟和噪声）的短音频剪辑，每个类别10个；测试数据包括一个2小时的声音文件和手动注释。数据通过自主录音单元收集，采样率为16 kHz，并经过带限能量检测和人工注释处理，适用于声音分类和迁移学习研究。