Automatic construction of accurate bioacoustics workflows under time constraints using a surrogate model

Automated bioacoustics analysis is being increasingly used to describe environmental phenomena such as species abundance and biodiversity. Within this research area, many algorithms have been proposed. These achieve different sub-objectives within bioacoustics processes and can be combined to form workflows. However, these algorithms are typically evaluated in a limited number of scenarios and are rarely evaluated with different combinations of other tasks. This can result in workflows that are not well optimised to serve a given scenario, particularly under resource and time constraints, which ultimately leads to inaccurate bioacoustics analyses. This work examines the problem of bioacoustics workflow construction by searching and ordering combinations of tasks to determine which produce the most accurate output while remaining under user-defined time constraints. Workflow construction is investigated within a scenario where species need to be classified within synthetically generated soundscapes with different numbers of species, noise levels, and densities of species. A search algorithm is created that applies Particle Swarm Optimisation (PSO) to a neural network-based surrogate model. This algorithm is used to efficiently search for candidate workflow structures. This is compared to a random search, a genetic algorithm, and a PSO algorithm without the surrogate model, as well as existing workflows based on previous research. It is found that for all scenarios, the surrogate model-based search method can quickly find effective workflows in a low number of searches. Furthermore, it is found that workflow effectiveness varies depending on the scenarios and recordings used.