Abstract: The crane(Grunidae) is not a songbird family, but it is known to communicate using calls such as unison calls. With the development of machine learning and deep learning algorithms, automated detection and classification of bird sounds using these techniques has been actively conducted, including the NMF algorithm. However, there is a lack of research on applying this method to the sounds of cranes and studying the sounds of wild individuals. Therefore, this study aims to classify the calls of the white-naped cranes(Grus vipio) that overwinter in the middle of the Korean peninsula and explore the frequency of occurrence according to the type of sound using the NMF-based algorithm among various machine learning algorithms.
The recorders are installed in a plain located in the Imjingang estuary in Paju, Gyeonggi-do. Data was recorded for 24 hours with a sample rate of 48000Hz. Until November 7th, recordings was recorded for 5 minutes with 15 minutes of rest, and after that, they were recorded for 1 minute with 9 minutes of rest. During the recordings in early November and early December, 510 syllables were used for analysis. The sounds of the white-naped cranes were classified into four types based on syllable duration and the number of pulses using Raven Pro software. For automated detection, the Soundscape_IR Python library based on the NMF algorithm was used to train five syllables per type from the entire dataset, and noise was removed using the Audacity program for all test files. Type-specific frequencies were calculated by comparing the annotated files based on syllable duration and syllable begin-time and end-time.
The results showed that type A had the highest frequency at 87.5%, followed by type 3 at 6.5%, type 2 at 4%, and type 4 at 2%. The detection rate was 60% comparing to the annotated data. It is because Gruidae is known to use different sound types in different situations. Therefore, further research is needed to analyze the correlation between syllable types and situations through observation studies. Moreover, the detection rate could be improved by adjusting parameters or using different algorithms to supplement the NMF method, which may not be suitable for certain types or in the presence of noise. Nevertheless, the study has significant implications as it reveals some regularities in sounds of wild individuals and shown a possibility of automated detection.
