生物多样性 ›› 2024, Vol. 32 ›› Issue (10): 24215. DOI: 10.17520/biods.2024215 cstr: 32101.14.biods.2024215
申小虎1,2,*()(
), 李冠宇3(
), 史洪飞2, 王传之4
收稿日期:
2024-06-01
接受日期:
2024-10-11
出版日期:
2024-10-20
发布日期:
2024-12-03
通讯作者:
*E-mail: shenxiaohu@jspi.cn
基金资助:
Xiaohu Shen1,2,*()(
), Guanyu Li3(
), Hongfei Shi2, Chuanzhi Wang4
Received:
2024-06-01
Accepted:
2024-10-11
Online:
2024-10-20
Published:
2024-12-03
Contact:
*E-mail: shenxiaohu@jspi.cn
Supported by:
摘要:
鸟声识别是被动声学监测的重要应用领域, 集成学习方法对提升鸟类识别精度具有重要研究价值, 但面对数据不平衡问题时缺少有效的集成策略。为此, 通过基学习器的迁移学习获得鸟声信号的不同方面表征, 满足了少标签样本条件下的学习训练。同时, 设计加入自注意力机制的特征融合和敏感正则项用于提升模型对稀有鸟类的关注度, 确保集成模型在信息不对称情况下推理时获得全局最优解。本文在南京老山森林公园共收集了10种鸟类样本, 并对预训练模型完成了微调。通过鸟声识别分类实验, 在样本不平衡的自建数据集与BirdCLEF 2023数据集上, 总体分类精度分别达到了95.29%和90.17%。本文所提出的集成学习策略提升了少量样本类别的敏感度, 增强了模型的泛化能力和学习训练效率, 与主流集成学习方法相比较, 能更好地适用于当地稀有鸟类的被动鸟声监测与识别, 助力鸟类生态环境的精准保护。
中图分类号:
申小虎, 李冠宇, 史洪飞, 王传之 (2024) 数据不平衡下鸟声识别的集成学习策略. 生物多样性, 32, 24215. DOI: 10.17520/biods.2024215.
Xiaohu Shen, Guanyu Li, Hongfei Shi, Chuanzhi Wang (2024) Ensemble learning strategy for birdsong recognition under data imbalance. Biodiversity Science, 32, 24215. DOI: 10.17520/biods.2024215.
基模型 Base_model | 输入 Input | 优点 Advantage | 缺点 Disadvantage | 特定问题 Specific issue |
---|---|---|---|---|
ConvNeXt_small | 固定尺寸的频谱图 Fixed size spectrogram | 在保持了卷积神经网络优点的同时, 借鉴了Transformer的先进设计理念, 提升了模型性能 While maintaining the advantages of CNN, the advanced design concept of Transformer has been referenced to improve model performance | 使用了大尺寸的卷积核, 计算量相对较大 Using a large-sized convolution kernel results in relatively high computational complexity | ‒ |
ECA-NFNet | 固定尺寸/可变尺寸的频 谱图 Fixed size/variable size spectrogram | 通过引入ECA模块, 提升了模型的注意力和特征提取能力, 使得模型在图像分类等任务上表现出色 By introducing the ECA module, the model’s attention and feature extraction capabilities were improved, resulting in excellent performance on image classification and other tasks | 模型可能较为复杂, 需要较多的计算资源 The model may be relatively complex and require substantial computational resources | ‒ |
EfficientNet-B0 | 固定尺寸的频谱图 Fixed size spectrogram | 在保持高性能的同时, 具有较小的计算成本和参数数量 While maintaining high performance, it possesses a smaller computational cost and fewer parameters | ‒ | 适应资源受限或需要高效推理的场景 Adapting to scenarios with limited resources or requiring efficient inference |
ReNeXt50 | 固定尺寸的频谱图 Fixed size spectrogram | 在保持参数和计算复杂度不变的情况下, 通过增加基数可以有效地提升模型性能 Increasing the base can effectively improve model performance while keeping parameters and computational complexity unchanged | 基数的增加可能会带来额外的计算负担 An increase in the base may introduce additional computational burden | |
EfficientNetV2 | 固定尺寸/可变尺寸的频 谱图 Fixed size/variable size spectrogram | 权重共享 + 自适应参数伸缩, 进一步提升效率 Weight sharing combined with adaptive parameter scaling further enhances efficiency | 更多的训练数据和计算资 源来达到最佳性能 Achieving optimal performance requires more training data and computational resources | 适应资源受限或需要高效推理的场景 Adapting to scenarios with limited resources or requiring efficient inference |
表1 主干网络基学习器的比较
Table 1 Comparison of base learners in backbone networks
基模型 Base_model | 输入 Input | 优点 Advantage | 缺点 Disadvantage | 特定问题 Specific issue |
---|---|---|---|---|
ConvNeXt_small | 固定尺寸的频谱图 Fixed size spectrogram | 在保持了卷积神经网络优点的同时, 借鉴了Transformer的先进设计理念, 提升了模型性能 While maintaining the advantages of CNN, the advanced design concept of Transformer has been referenced to improve model performance | 使用了大尺寸的卷积核, 计算量相对较大 Using a large-sized convolution kernel results in relatively high computational complexity | ‒ |
ECA-NFNet | 固定尺寸/可变尺寸的频 谱图 Fixed size/variable size spectrogram | 通过引入ECA模块, 提升了模型的注意力和特征提取能力, 使得模型在图像分类等任务上表现出色 By introducing the ECA module, the model’s attention and feature extraction capabilities were improved, resulting in excellent performance on image classification and other tasks | 模型可能较为复杂, 需要较多的计算资源 The model may be relatively complex and require substantial computational resources | ‒ |
EfficientNet-B0 | 固定尺寸的频谱图 Fixed size spectrogram | 在保持高性能的同时, 具有较小的计算成本和参数数量 While maintaining high performance, it possesses a smaller computational cost and fewer parameters | ‒ | 适应资源受限或需要高效推理的场景 Adapting to scenarios with limited resources or requiring efficient inference |
ReNeXt50 | 固定尺寸的频谱图 Fixed size spectrogram | 在保持参数和计算复杂度不变的情况下, 通过增加基数可以有效地提升模型性能 Increasing the base can effectively improve model performance while keeping parameters and computational complexity unchanged | 基数的增加可能会带来额外的计算负担 An increase in the base may introduce additional computational burden | |
EfficientNetV2 | 固定尺寸/可变尺寸的频 谱图 Fixed size/variable size spectrogram | 权重共享 + 自适应参数伸缩, 进一步提升效率 Weight sharing combined with adaptive parameter scaling further enhances efficiency | 更多的训练数据和计算资 源来达到最佳性能 Achieving optimal performance requires more training data and computational resources | 适应资源受限或需要高效推理的场景 Adapting to scenarios with limited resources or requiring efficient inference |
鸟声种类 Bird species | 科 Family | 物种丰度 Species richness | 样本时长 Sample duration (s) | 鸟声样本数 Number of birdsong samples | |
---|---|---|---|---|---|
数据集1 Dataset 1 | 数据集2 Dataset 2 | ||||
灰喜鹊 Cyanopica cyanus | 鸦科 Corvidae | 常见 Common | 3,104 | 621 | 621 |
环颈鸻 Charadrius alexandrinus | 鸻科 Charadriidae | 少见 Rare | 765 | 153 | 47 |
白腰文鸟 Lonchura striata | 麻雀科 Passeridae | 极少见 Extremely rare | 277 | 56 | 19 |
树麻雀 Passer montanus | 麻雀科 Passeridae | 常见 Common | 5,106 | 1,021 | 1,021 |
雉鸡 Phasianus colchicus | 雉科 Phasianidae | 较常见 Relatively common | 1,718 | 343 | 160 |
山斑鸠 Streptopelia orientalis | 鸠鸽科 Columbidae | 常见 Common | 2,447 | 489 | 489 |
黄腰柳莺 Phylloscopus proregulus | 莺科 Sylviidae | 少见 Rare | 452 | 90 | 30 |
八哥 Acridotheres cristatellus | 椋鸟科 Sturnidae | 常见 Common | 2,709 | 541 | 541 |
银喉长尾山雀 Aegithalos glaucogularis | 山雀科 Paridae | 较常见 Relatively common | 1,712 | 342 | 342 |
赤腹鹰 Accipiter soloensis | 鹰科 Accipitridae | 较常见 Relatively common | 814 | 163 | 814 |
表2 实验中自建的南京老山地区鸟声数据集
Table 2 The self-built birdsong dataset from Laoshan area, Nanjing used in the experiment
鸟声种类 Bird species | 科 Family | 物种丰度 Species richness | 样本时长 Sample duration (s) | 鸟声样本数 Number of birdsong samples | |
---|---|---|---|---|---|
数据集1 Dataset 1 | 数据集2 Dataset 2 | ||||
灰喜鹊 Cyanopica cyanus | 鸦科 Corvidae | 常见 Common | 3,104 | 621 | 621 |
环颈鸻 Charadrius alexandrinus | 鸻科 Charadriidae | 少见 Rare | 765 | 153 | 47 |
白腰文鸟 Lonchura striata | 麻雀科 Passeridae | 极少见 Extremely rare | 277 | 56 | 19 |
树麻雀 Passer montanus | 麻雀科 Passeridae | 常见 Common | 5,106 | 1,021 | 1,021 |
雉鸡 Phasianus colchicus | 雉科 Phasianidae | 较常见 Relatively common | 1,718 | 343 | 160 |
山斑鸠 Streptopelia orientalis | 鸠鸽科 Columbidae | 常见 Common | 2,447 | 489 | 489 |
黄腰柳莺 Phylloscopus proregulus | 莺科 Sylviidae | 少见 Rare | 452 | 90 | 30 |
八哥 Acridotheres cristatellus | 椋鸟科 Sturnidae | 常见 Common | 2,709 | 541 | 541 |
银喉长尾山雀 Aegithalos glaucogularis | 山雀科 Paridae | 较常见 Relatively common | 1,712 | 342 | 342 |
赤腹鹰 Accipiter soloensis | 鹰科 Accipitridae | 较常见 Relatively common | 814 | 163 | 814 |
集成方案 Integration solution | 基学习器 Base learner | 识别敏感精度 Sensitive mean average precision (%) | ||||
---|---|---|---|---|---|---|
EfficientNet-B0 | ConvNeXt | ECA-NFNet | ReNeXt50 | EfficientNet V2 | 老山数据集1 Laoshan dataset 1 | |
方案1 Option 1 | √ | - | - | - | - | 90.73 |
方案2 Option 2 | - | √ | - | - | - | 92.28 |
方案3 Option 3 | - | - | √ | - | - | 89.70 |
方案4 Option 4 | - | - | - | √ | - | 88.22 |
方案5 Option 5 | - | - | - | - | √ | 91.49 |
方案6 Option 6 | √ | - | - | - | √ | 93.19 |
方案7 Option 7 | - | - | √ | - | √ | 92.41 |
方案8 Option 8 | - | √ | √ | - | - | 94.24 |
方案9 Option 9 | √ | - | - | √ | √ | 94.50 |
方案10 Option 10 | √ | - | √ | √ | √ | 93.85 |
方案11 Option 11 | √ | √ | √ | √ | √ | 95.29 |
表3 基学习器集成方案对比
Table 3 Comparison of basic learner integration methods
集成方案 Integration solution | 基学习器 Base learner | 识别敏感精度 Sensitive mean average precision (%) | ||||
---|---|---|---|---|---|---|
EfficientNet-B0 | ConvNeXt | ECA-NFNet | ReNeXt50 | EfficientNet V2 | 老山数据集1 Laoshan dataset 1 | |
方案1 Option 1 | √ | - | - | - | - | 90.73 |
方案2 Option 2 | - | √ | - | - | - | 92.28 |
方案3 Option 3 | - | - | √ | - | - | 89.70 |
方案4 Option 4 | - | - | - | √ | - | 88.22 |
方案5 Option 5 | - | - | - | - | √ | 91.49 |
方案6 Option 6 | √ | - | - | - | √ | 93.19 |
方案7 Option 7 | - | - | √ | - | √ | 92.41 |
方案8 Option 8 | - | √ | √ | - | - | 94.24 |
方案9 Option 9 | √ | - | - | √ | √ | 94.50 |
方案10 Option 10 | √ | - | √ | √ | √ | 93.85 |
方案11 Option 11 | √ | √ | √ | √ | √ | 95.29 |
图7 本文方法与不同消融模型下的曲线下面积对比。CSE-BSR: 改进代价敏感集成学习策略; CSE-BSR-RPB: 上下文建模与特征变换消融模型; CSE-BSR-CMT: 稀有惩罚偏置消融模型。
Fig. 7 Comparison of the area under the curve (AUC) between the proposed method and different ablation models. CSE-BSR, Cost-sensitive stacking ensemble for bird sound recognition; CSE-BSR-RPB, Ablation model for context modeling and transform; CSE-BSR-CMT, Ablation model for rarity penalty bias.
方法 Method | 融合方式 Fusion method | 集成权重计算 Integrated weight calculation | 识别敏感精度 Sensitive mean average precision (%) | ||
---|---|---|---|---|---|
老山数据集1 Laoshan dataset 1 | 老山数据集2 Laoshan dataset 2 | 公开数据集 Open dataset | |||
ResNet-34 + EfficientNetV2 (2021年挑战赛亚军 The 2nd place in the 2021 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 93.11 | 91.08 | 88.57 |
EfficientNet-B3 + ECA-NFNet (2022年挑战赛冠军 The 1st place in the 2022 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 92.47 | 90.40 | 87.65 |
ResNet-152 + Inception (2019年挑战赛冠军 The 1st place in the 2019 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 90.60 | 90.85 | 86.16 |
ECA-NFNet + ConvNeXt + ConvNeXt V2 (2023年挑战赛冠军 The 1st place in the 2023 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 91.62 | 91.49 | 89.10 |
EfficientNetV2 + ResNet-34 + EfficientNet-B0 + EfficientNet-B3 (2023年挑战赛亚军 The 2nd place in the 2023 challenge) | 决策融合 Decision fusion | 排名加权 Ranking weighting | 94.64 | 92.76 | 89.52 |
EfficientNet-B0 + ReNeXt50 + SeResNet50 + DenseNet121 + ReNeXt50 + ResNeSt121 (Conde et al, | 特征融合 Feature fusion | 简单平均 Simple average | 93.55 | 92.23 | 89.91 |
EfficientNet-B0 + ConvNeXt_small + ECA- NFNet + ReNeXt50 + EfficientNet V2 (本文方法 The method presented in this paper) | 特征融合 Feature fusion | 自注意力加权 Self-attention weighting | 95.29 | 94.58 | 90.17 |
表4 本文方法与最新方法的识别敏感精度对比
Table 4 Comparison of sensitive mean average precision between the proposed method and state-of-the-art methods
方法 Method | 融合方式 Fusion method | 集成权重计算 Integrated weight calculation | 识别敏感精度 Sensitive mean average precision (%) | ||
---|---|---|---|---|---|
老山数据集1 Laoshan dataset 1 | 老山数据集2 Laoshan dataset 2 | 公开数据集 Open dataset | |||
ResNet-34 + EfficientNetV2 (2021年挑战赛亚军 The 2nd place in the 2021 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 93.11 | 91.08 | 88.57 |
EfficientNet-B3 + ECA-NFNet (2022年挑战赛冠军 The 1st place in the 2022 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 92.47 | 90.40 | 87.65 |
ResNet-152 + Inception (2019年挑战赛冠军 The 1st place in the 2019 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 90.60 | 90.85 | 86.16 |
ECA-NFNet + ConvNeXt + ConvNeXt V2 (2023年挑战赛冠军 The 1st place in the 2023 challenge) | 决策融合 Decision fusion | 简单平均 Simple average | 91.62 | 91.49 | 89.10 |
EfficientNetV2 + ResNet-34 + EfficientNet-B0 + EfficientNet-B3 (2023年挑战赛亚军 The 2nd place in the 2023 challenge) | 决策融合 Decision fusion | 排名加权 Ranking weighting | 94.64 | 92.76 | 89.52 |
EfficientNet-B0 + ReNeXt50 + SeResNet50 + DenseNet121 + ReNeXt50 + ResNeSt121 (Conde et al, | 特征融合 Feature fusion | 简单平均 Simple average | 93.55 | 92.23 | 89.91 |
EfficientNet-B0 + ConvNeXt_small + ECA- NFNet + ReNeXt50 + EfficientNet V2 (本文方法 The method presented in this paper) | 特征融合 Feature fusion | 自注意力加权 Self-attention weighting | 95.29 | 94.58 | 90.17 |
[1] | Acconcjaioco M, Ntalampiras S (2021) One-shot learning for acoustic identification of bird species in non-stationary environments. In: 25th International Conference on Pattern Recognition, pp. 755-762. Milan, Italy. |
[2] | Brock A, De S, Smith SL (2021) High-performance large-scale image recognition without normalization. In: 38th International Conference on Machine Learning, pp. 1059-1071. |
[3] | Chen JP, Lin H, Wu XF, Huang YD, Cheng JJ, Zhuang YB (2023) Classification forecast method of costal low visibility weather based on ensemble learning. Journal of Tropical Meteorology, 39, 680-688. (in Chinese with English abstract) |
[陈锦鹏, 林辉, 吴雪菲, 黄奕丹, 程晶晶, 庄毅斌 (2023) 基于集成学习的沿海低能见度天气分类预报方法. 热带气象学报, 39, 680-688.] | |
[4] | Chen TY, Zhang XR, Yang ZX, Zhang F, Xu P, Lu CH (2020) Composition and inter-annual dynamics of the bird community during breeding season at Laoshan National Forest Park, Nanjing. Chinese Journal of Wildlife, 41, 339-348. (in Chinese with English abstract) |
[陈泰宇, 张啸然, 杨再玺, 张芳, 许鹏, 鲁长虎 (2020) 南京老山国家森林公园繁殖期鸟类组成及年间变化. 野生动物学报, 41, 339-348.] | |
[5] | Conde MV, Shubham K, Agnihotri P, Movva ND, Bessenyei S (2021) Weakly-supervised classification and detection of bird sounds in the wild. A BirdCLEF 2021 solution. arXiv: 2107.04878. |
[6] | El-Rashidy N, El-Sappagh S, Abuhmed T, Abdelrazek S, El-Bakry HM (2020) Intensive care unit mortality prediction: An improved patient-specific stacking ensemble model. IEEE Access, 8, 133541-133564. |
[7] | Hu GN, Wang DL (2010) A Tandem Algorithm for pitch estimation and voiced speech segregation. IEEE Transactions on Audio, Speech, and Language Processing, 18, 2067-2079. |
[8] | Huang G, Liu Z, Laurens VDM, Weinberger KQ (2017) Densely connected convolutional networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, pp. 2261-2269. Honolulu, USA. |
[9] |
Ke PF, Cai MG, Wu T (2020) Face recognition algorithm based on improved convolutional neural network and ensemble learning. Computer Engineering, 46, 262-267, 273. (in Chinese with English abstract)
DOI |
[柯鹏飞, 蔡茂国, 吴涛 (2020) 基于改进卷积神经网络与集成学习的人脸识别算法. 计算机工程, 46, 262-267, 273.]
DOI |
|
[10] | Kong DM, Chen HJ, Chen XY, Dong R, Wang ST (2021) Research on oil identification method based on three- dimensional fluorescence spectroscopy combined with sparse principal component analysis and support vector machine. Spectroscopy and Spectral Analysis, 41, 3474-3479. (in Chinese with English abstract) |
[孔德明, 陈红杰, 陈晓玉, 董瑞, 王书涛 (2021) 三维荧光光谱结合稀疏主成分分析和支持向量机的油类识别方法研究. 光谱学与光谱分析, 41, 3474-3479.] | |
[11] | Lasseck M (2019) Bird species identification in soundscapes. In:Conference and Labs of the Evaluation Forum 2019. Lugano, Switzerland. |
[12] | Li LH, Dong HY, Liu WJ, Li BL, Dai Q (2024) Neighborhood-tolerance mutual information selection ensemble classification algorithm for incomplete data sets. Journal of Nanjing University (Natural Science), 60, 106-117. (in Chinese with English abstract) |
[李丽红, 董红瑶, 刘文杰, 李宝霖, 代琪 (2024) 不完备数据集的邻域容差互信息选择集成分类算法. 南京大学学报(自然科学), 60, 106-117.] | |
[13] | Lin TY, Goyal P, Girshick R, He KM, Dollár P (2020) Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42, 318-327. |
[14] | Liu Z, Mao HZ, Wu CY, Feichtenhofer C, Darrell T, Xie S (2022) A ConvNet for the 2020s. arXiv: 2201.03545. |
[15] | Luo CW, Wang SS, Yin JS, Zhu SY, Lin B, Cao J (2023) Research status and prospect of ensemble learning. Journal of Command and Control, 9, 1-8. (in Chinese with English abstract) |
[罗常伟, 王双双, 尹峻松, 朱思宇, 林波, 曹江 (2023) 集成学习研究现状及展望. 指挥与控制学报, 9, 1-8.] | |
[16] |
Shen XH, Zhu XY, Shi HF, Wang CZ (2023) Research progress of birdsong recognition algorithms based on machine learning. Biodiversity Science, 31, 23272. (in Chinese with English abstract)
DOI |
[申小虎, 朱翔宇, 史洪飞, 王传之 (2023) 基于机器学习鸟声识别算法研究进展. 生物多样性, 31, 23272.]
DOI |
|
[17] | Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z (2016) Rethinking the inception architecture for computer vision. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818-2826. Las Vegas, USA. |
[18] | Tan MX, Le QV (2019) EfficientNet: Rethinking model scaling for convolutional neural networks. arXiv: 1905.11946. |
[19] | Tan MX, Le QV (2021) EfficientNetV2: Smaller models and faster training. arXiv: 2104.00298. |
[20] | Wang JD, Zhang D, Zhou Y (2022) Ensemble deep learning for automated classification of power quality disturbances signals. Electric Power Systems Research, 213, 108695. |
[21] | Wang QL, Wu BG, Zhu PF, Li PH, Zuo WM, Hu QH (2020) ECA-Net:Efficient channel attention for deep convolutional neural networks. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11531-11539. Seattle, USA. |
[22] | Xie SN, Girshick R, Dollár P, Tu ZW, He KM (2017) Aggregated residual transformations for deep neural networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, pp. 5987-5995. Honolulu, USA. |
[23] | Zhang FY, Zhang LY, Chen HX, Xie JJ (2021) Bird species identification using spectrogram based on multi-channel fusion of DCNNs. Entropy, 23, 1507-1518. |
[24] | Zhou T, Dong YL, Huo BQ, Liu S, Ma ZJ (2021) U-Net and its applications in medical image segmentation: A review. Journal of Image and Graphics, 26, 2058-2077. (in Chinese with English abstract) |
[周涛, 董雅丽, 霍兵强, 刘珊, 马宗军 (2021) U-Net网络医学图像分割应用综述. 中国图象图形学报, 26, 2058-2077.] | |
[25] | Zhu ZW, Lan X, Zhao TT, Guo YM, Kojodjojo P, Xu ZY, Liu Z, Liu SQ, Wang H, Sun XZ, Feng ML (2021) Identification of 27 abnormalities from multi-lead ECG signals: An ensembled SE_ResNet framework with sign loss function. Physiological Measurement, 42, 065008. |
[1] | 郭倩茸, 段淑斐, 谢捷, 董雪燕, 肖治术. 鸟声标注技术及其在被动声学监测中的应用[J]. 生物多样性, 2024, 32(10): 24313-. |
[2] | 黄万涛, 郝泽周, 张梓欣, 肖治术, 张承云. 被动声学监测设备性能比较及对鸟声识别的影响[J]. 生物多样性, 2024, 32(10): 24273-. |
[3] | 申小虎, 朱翔宇, 史洪飞, 王传之. 基于机器学习鸟声识别算法研究进展[J]. 生物多样性, 2023, 31(11): 23272-. |
阅读次数 | ||||||
全文 |
|
|||||
摘要 |
|
|||||
备案号:京ICP备16067583号-7
Copyright © 2022 版权所有 《生物多样性》编辑部
地址: 北京香山南辛村20号, 邮编:100093
电话: 010-62836137, 62836665 E-mail: biodiversity@ibcas.ac.cn