machinelearningIt is important to gain a better understanding about the climate and ecological changes in the world. One way to address this is to study seasonal migration patterns in songbird populations, since birds respond quickly to environmental changes. During migratory periods, many species of songbirds use flight calls, which are species-specific and are distinct from other vocalizations. Therefore, flight calls information can be used to determine the relative abundance of species and is important to understand long-term population trends. Due to costly human effort to collect data about birds in traditional methods, using machine learning (ML) methods to identify bird species from continuous audio recordings has been a hot topic in in recent conference competitions. Although there are some recent advances it is still an open ML problem to reliably identify bird sounds in field recordings data due to simultaneously vocalizing birds and various background noise.

In this project, we focus on critical aspects of this problem. The total process is divided into four steps. First, Audio data are preprocessed into spectrograms, which are further cleaned by applying background noise reduction and image processing techniques, and connected pixels (acoustic patterns) in the spectrograms are labeled into rectangle segments. Second, features are then extracted and selected from different sources, e.g., file statistics, segment statistics and probabilities, and mel-frequency cepstral coefficients (MFCCs). Third, the classification is then done by using multiple algorithms, e.g., naive Bayes, k-nearest neighbors (k-NN), support vector machines (SVM), etc. Finally, we will explore some ensemble methods for further exploring some properties on overall performance by combining the predictions of models, as well as facilitating scalability in real-world usages.

The software developed for this project will be used by the Carnegie Museum of Natural History, and possibly shared with other land managers, researchers, and educators to enhance the use of flight calls as a method to study the patterns of migratory songbirds.