关键词: 鸟类, 声纹识别, 中国鸟类数据库, 空间分布模型, 联合模型

AbstractBirds are key indicator taxa for biodiversity conservation due to their sensitive responses to habitat changes. The development of bird vocalization recognition models in China is constrained by incomplete datasets for model training, insufficient integration of geographical distribution information, and a lack of field validation through comparative studies. To address these gaps, this study constructed an acoustic dataset comprising vocalizations of 1,368 bird species in China. A bird vocalization recognition model was developed using a two-stage transfer learning approach. Concurrently, a species spatial distribution probability model at the municipal scale, covering 1,428 species, was established based on citizen science bird observation data. Finally, a joint bird vocalization recognition model integrating acoustic and spatial distribution information via a Sigmoid function was developed and applied. Model performance was validated using a dataset of vocalizations from 66 common nationwide bird species. The results indicate that as the confidence threshold increases, the joint model demonstrates more stable performance than the standalone acoustic recognition model, with a maximum accuracy improvement of 16.8%. For practical application, a comparative analysis is conducted using four years of data from site-based monitoring and manual synchronous observation. In terms of species richness, the machine identifies 68 species, accounting for 60.2% of the species recorded manually. The machine-based method reflects the activity patterns of bird species accurately and responds more rapidly to changes in bird activity rhythms.

Key words: Bird vocalization, acoustic recognition, China bird database, species distribution model, joint model