Abstract: Information on species identity, abundance, and habitat is the fundamental requirement in bird ecology research. The rapid development of artificial intelligence technology based on deep learning algorithm has made up for the shortcomings of traditional field investigation methods, providing an intelligent means for bird ecology research. By building a ResNet34 dualtask network model with adaptive loss function, a bird and habitat recognition model was built to realize the simultaneous identification of bird species and habitat types across four observation sites in Beijing Cuihu National Urban Wetland Park. A total of 10 bird species were recorded based on this model, including swan goose, gray heron, great cormorant and mallard. Night heron and gray heron use trees as main habitats. Swan goose and mallard use water as main habitats. Great cormorant, eastern spot-billed duck, and little egret use artificial habitats such as standing trees in water as main habitats. Common magpies mainly use the ground as habitat. Ruddy shelduck and mandarin duck have a wide range of habitats, which are distributed on the ground, artificial habitats, and water. The recognition accuracy of the proposed model reaches 95.62% for bird species and 97.20% for habitat types. The bird image acquisition method based on deep learning technology was used to replace the artificial data collection means, and the dual-task branch structure of “species + habitat” was used for the first time to identify species and habitat, which greatly improved the efficiency and ensures the objectivity and accuracy of data collection. It realizes the effective combination of avian ecology research and artificial intelligence, which has reference significance for the evolution of ecological research methods.

Key words: artificial intelligence, deep learning, avian diversity, habitat