Abstract
Birds serve as valuable indicators for assessing biodiversity. Especially in Europe and the United States, acoustic monitoring using machine learning technologies is getting attention to automatically identify species to monitor birds efficiently. However, acoustic monitoring studies focused on Japanese birds are few. Moreover, most of the studies attempt to identify only one or a limited number of bird species. Multi-bird species identification systems are expected to enhance the applicability of bird monitoring in Japan. This paper describes a multi-species identification system for Japanese birds that we developed. First, we developed an identification system for 66-bird species using the Audio Spectrogram Transformer, an attention-based model for audio identification. The system showed an overall accuracy of 0.801, a precision of 0.797, a recall of 0.780, and an F1 score of 0.782. We also tested the performance for classifying unknown data. The test results yielded an overall accuracy of 0.677, a precision of 0.679, a recall of 0.635, and an F1 score of 0.670, surpassing previous research performance. However, some species yielded lower accuracy, possibly due to insufficient training data and bird song variations caused by differences in geographical area, individuals, recording equipment, or recording conditions.
Additionally, we discussed future perspectives for the social implementation of bird species identification systems. Our findings emphasized the advantage of utilizing citizen science data and selecting appropriate techniques depending on specific purposes in system development. Proper research design using the advantages and disadvantages identified from the results of the overall workflow was also highlighted as crucial for future advancements.
