Enhancing Bird Migration Studies: Detecting Birdsong in Audio Files Using Convolutional Neural Networks

Monitoring climate threats and global environmental changes is a crucial issue worldwide. To achieve this, it is important to continuously study these changes using non-standard approaches. One such approach is studying bird migration. It is crucial for analyzing environmental factors that cover large areas of land, such as temperature, air pollution, and water pollution. This article presents a machine-learning model that accurately identifies birdsong in audio files to study bird migration in specific areas. This model uses a convolutional neural network to classify audio files based on images of their spectrograms. The study involves collecting and analyzing audio files that can be used to identify characteristics according to which the sound in the file will be identified as bird sound. The study investigated the effectiveness of using STFT and Mel-spectrograms in a convolutional neural network to classify files based on the presence of bird sounds. The model trained on Mel spectrograms achieved an accuracy of 80%, which is 17% higher than the model trained on STFT spectrograms. This accuracy value is completely sufficient to claim the presence of bird sounds in the audio file. Particular attention is given to the efficiency and accuracy of the CNN model because these indicators enable the comparison and selection of the best classifier for a given file type and model. Although the classification accuracy of the chosen method was lower than expected when building the classification model, the model can predict an audio file in less than 0,5 s, which is a significant achievement. The obtained results may aid in the development of systems for monitoring bird migration across vast territories. To enhance the technique, multiple sets of audio files can be utilized to classify the presence of birds in the environment, improving the accuracy of the method and model. However, the simultaneous complex processing of numerous audio files may reduce processing speed. This study will be useful for combining methods of tracking birds in the environment, such as a combination of direct observation and auditory methods.