文献类型：期刊文献

中文题名：被动声学监测设备性能比较及对鸟声识别的影响

英文题名：A comparison of bird sound recognition performance among acoustic recorders

作者：黄万涛[1] 郝泽周[2] 张梓欣[1] 肖治术[3] 张承云[1]

第一作者：黄万涛

机构：[1]广州大学电子与通信工程学院,广州510006;[2]中国林业科学研究院热带林业研究所,广州510520;[3]中国科学院动物研究所农业虫害鼠害综合治理研究国家重点实验室,北京100101

年份：2024

卷号：32

期号：10

起止页码：94-103

中文期刊名：生物多样性

外文期刊名：Biodiversity Science

收录：;北大核心:【北大核心2023】；CSCD:【CSCD2023_2024】；

基金：国家自然科学基金(32171520)。

语种：中文

中文关键词：被动声学监测;声学录音设备;鸟声识别;深度学习;BirdNET

外文关键词：passive acoustic monitoring;acoustic recording devices;bird sound recognition;deep learning;BirdNET

分类号：TN9

摘要：被动声学监测技术能够以非侵入的方式进行长期有效的监测,已广泛应用于鸟类的监测,监测过程收集到的大量数据需要借助自动化识别技术进行分析处理。然而,不同录音设备的性能差异可能会影响自动化识别软件正确识别鸟类类别的能力。本研究使用国内外6种类型录音设备对4种不同频带范围的鸟声信号进行回放录音,选取BirdNET作为鸟类鸣声自动识别器,对2种植被类型录音环境、5种距离和3种声源方向的回放录音信号进行鸟声识别,评估这些变量对鸟类类别识别性能的影响。通过比较录音设备的基本参数和配置,并构建广义线性模型(generalized linear model,GLM)对识别结果进行统计分析,以评估不同录音设备的监测性能。结果表明,录音设备类型显著影响BirdNET对鸟类类别的识别准确率。总体上,随着距离增加,设备的监测有效性下降,且在50 m或更近距离内,BirdNET的识别准确率显著更高。声源方向对识别性能也有影响,当声源与录音设备方向相反时,识别准确率显著下降。不同设备对4种不同频带范围鸟声信号的识别有效性存在不一致性。此外,植被类型显著影响鸟声信号传播的衰减,草地植被下的总体识别准确率比林地植被高40.1%。本研究建议,在选择和部署长期录音监测设备前,除评估成本和参数外,还应进行实地录音监测有效性的评估。根据评估结果,优化监测距离和方向设置,以提升监测策略的有效性。
Aims:Passive acoustic monitoring technology has been widely used for monitoring bird species,enabling non-invasive and long-term effective monitoring.Extensive data collection requires automated identification technologies for effective analysis.However,differences in recording device performance can affect the accuracy of automated software in identifying bird species.Methods:Six separate recording devices from various manufacturers are tested by recording bird call playback across four frequency bands.We use BirdNET as the automatic bird sound identifier under two types of vegetation environment,five categories of distance between the recording devices and sound source,and three sound source directions.Our goal is evaluating the impact of these variables on bird species identification performance.We assess the monitoring performance of different recording devices by comparing the basic parameters and configurations of the devices and constructing a generalized linear model(GLM)to statistically analyze the identification results.Results:Our analysis suggests the type of recording device significantly affects the ability for BirdNET to correctly identify bird species.As distance increases,the effectiveness of the devices in monitoring decreases,with the identification accuracy of BirdNET significantly higher for distances within 50 meters than beyond.Further,the direction of sound impacts identification performance,with accuracy significantly decreasing when the sound source is in opposite direction of the recording device in identifying the four types of bird sound signals with different frequency bandwidth ranges.Additionally,the vegetation type significantly affects the attenuation of bird call signals,with overall identification accuracy in grassland vegetation 40.1%higher than forest vegetation.Conclusions:Our findings suggest the effectiveness of field recording monitoring should be assessed before selecting and deploying long-term recording monitoring equipment,in addition to evaluating equipment costs and parameters.Based on our evaluation,monitoring distance and direction settings should be optimized to enhance the effectiveness of monitoring strategies.