融合音频及生态位信息的跨地域鸟类物种识别方法  

Cross-regional bird species recognition method integrating audio and ecological niche information

作　　者：谢将剑[1,2] 沈忱 张飞宇 肖治术 Jiangjian Xie;Chen Shen;Feiyu Zhang;Zhishu Xiao(School of Technology,Beijing Forestry University,Beijing 100083,China;State Key Laboratory of Efficient Production of Forest Resources,Beijing 100083,China;State Key Laboratory of Integrated Management of Pest Insects and Rodents in Agriculture,Institute of Zoology,Chinese Academy of Sciences,Beijing 100101,China)

机构地区：[1]北京林业大学工学院,北京100083 [2]林木资源高效生产全国重点实验室,北京100083 [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

出　　处：《生物多样性》2024年第10期104-113,共10页Biodiversity Science

基　　金：国家自然科学基金(62303063);国家重点研发计划(E31OP01233)。

摘　　要：鸟类被动声学监测对于了解其种群和群落动态及相关物种的行为功能具有重要意义。利用深度学习技术和鸟类鸣声特征来自动识别鸟类物种,是实现鸟类大规模被动声学监测的关键。亲缘关系相近的鸟类物种的鸣声极为相似,容易混淆,使得假阳性增加,从而导致深度学习模型识别精度有限。针对上述问题,本文提出一种融合音频及生态位信息的鸟类物种识别方法:首先基于残差网络ResNet18构建音频识别模型,再使用最大熵模型对鸟类物种分布进行预测,获取鸟类物种在不同位置的适生指数作为生态位信息,最后构建融合音频及生态位信息的鸟类物种识别模型NicheNet。实验结果表明,与ResNet18相比,NicheNet的Top-1准确率提升了12.9%,Top-5准确率提升了10.6%,同时NicheNet的近种错误率、近属错误率以及近科错误率分别下降了3.1%、1.8%以及8.0%。结合对两对鸣声相似的鸟类同科物种的识别结果发现,NicheNet能够根据生态位信息对基于音频特征的分类结果进行修正,以提高对亲缘关系相近、鸣声相似而分布差异大的鸟类同科物种的识别效果。本文所提出的融合音频及生态位信息的鸟类物种识别方法能够有效降低亲缘关系相近、鸣声相似但生态位不同的鸟类物种的误识别率,进而提升基于鸣声的跨地域鸟类物种识别准确率。Aim:Passive acoustic monitoring plays a pivotal role in studying avian populations,community dynamics,and behaviors.For extensive passive acoustic monitoring,employing deep learning techniques to automatically identify bird species from their vocalizations is essential.However,closely related species often produce highly similar calls,leading to confusion and false positives,which can compromise the effectiveness of deep learning models.This paper presents a novel method that integrates audio data with ecological niche information to enhance species recognition accuracy.Here,‘ecological niche’encompasses a species’role in its environment,including its habitat,diet,and behavior.Methods:The approach begins with the development of an audio recognition model using the ResNet18,a prominent deep learning framework known for its capability to extract high-level features from audio signals.Subsequently,a maximum entropy model is employed to estimate the distribution of bird species and derive ecological suitability indices for various locations.These indices provide the necessary ecological niche information.An integrated model,NicheNet,is then constructed to combine audio features with ecological niche data for improved species recognition.Results&Conclusion:The integration of audio and ecological niche information through NicheNet demonstrates substantial improvements in recognition accuracy.Specifically,NicheNet enhances Top-1 recognition accuracy by 12.9% and Top-5 recognition accuracy by 10.6% compared to the ResNet18 model.Additionally,NicheNet reduces the near species error rate by 3.1%,the near genus error rate by 1.8%,and the near family error rate by 8.0%.Analysis of recognition outcomes for congeneric species with similar vocalizations reveals that NicheNet significantly refines classification by leveraging ecological niche information,thereby improving the discrimination of vocally similar but ecologically distinct species.This method effectively addresses the challenge of misidentification among closel

关 键 词：被动声学监测 鸟类鸣声识别 残差网络 深度学习 生态位 

分 类 号：Q958[生物学—动物学]