Bats inhabit all continents except Antarctica, and they have enormous potential as bioindicators. Therefore, monitoring bats helps us to understand the surrounding environmental changes. However, bats are nocturnal, which makes it difficult to visually monitor their behavior. This paper proposes a bat species identifier method based on the analysis of ultrasound called echolocation calls, which is a promising method to monitor bats' activity levels effectively. We develop a robust method to identify the bat species with improved accuracy by analyzing their echolocation calls. First, 1400 sound files with four families, 13 genera, and 30 species were recorded in Japan and the Jincheon-gun in South Korea from 1999 to 2019. Bat echolocation calls were detected from the sound files and used to generate 54,525 spectrograms by applying short-time Fourier transform. We developed a deep learning–based bat species identifier using convolutional neural networks with MobileNetV1 used as the model's architecture. Furthermore, we applied nested cross-validation with the Bayesian optimization algorithm to search for the optimal combination of hyperparameters and evaluate the expected performance. We achieved 98.1% accuracy, which outperformed previous studies that treated more than 30 bat species. We visualized important regions of the spectrograms which correspond to prediction using the Guided Grad-CAM. Moreover, we discussed how to treat the noise class and minimize the model training time. Then, we proposed potential solutions to boost the identifier's performance, the generalization of the echolocation call recording protocols, and applicable techniques to improve the identification accuracy. Future perspectives are 1) to change the deep learning algorithm from image classification to object detection and 2) to apply the proposed identifier to unknown bat echolocation calls to evaluate the feasibility of estimating bat fauna and spatial activity distribution.

蝙蝠居住在除南极洲以外的所有大洲，它们具有巨大的生物指示剂潜力。因此，监视蝙蝠有助于我们了解周围的环境变化。但是，蝙蝠是夜行性的，因此很难从视觉上监视它们的行为。本文提出了一种基于超声分析的蝙蝠种类识别方法，称为回声定位调用，这是一种有效监测蝙蝠活动水平的有前途的方法。我们开发了一种可靠的方法，通过分析回声定位信号来识别蝙蝠物种，从而提高准确性。首先，从1999年到2019年，在日本和韩国的镇川郡记录了1400个声音文件，这些文件有四个科，13个属和30个物种。从声音文件中检测到蝙蝠回声定位调用，并生成了54个声音文件。通过应用短时傅立叶变换获得525个频谱图。我们使用卷积神经网络开发了基于深度学习的蝙蝠物种识别器，并将MobileNetV1用作模型的体系结构。此外，我们将贝叶斯优化算法与嵌套交叉验证结合使用，以搜索超参数的最佳组合并评估预期性能。我们达到了98.1％的准确度，胜过之前对30多种蝙蝠种类进行研究的研究。我们可视化了频谱图的重要区域，这些区域对应于使用Guided Grad-CAM进行的预测。此外，我们讨论了如何处理噪声等级并最大程度地减少模型训练时间。然后，我们提出了可能的解决方案来提高标识符的性能，echolocation呼叫记录协议的一般化，以及提高识别准确性的适用技术。未来的观点是：1）将深度学习算法从图像分类更改为对象检测； 2）将拟议的标识符应用于未知的蝙蝠回声定位调用，以评估估计蝙蝠动物群和空间活动分布的可行性。