LUdev logo
Back to portfolio

Classification of Insect Sounds

1 / 2

The project

Our project aimed to develop a reliable neural network to non-invasively monitor insect populations. Two model architectures were built, both sharing a learnable audio frontend (LEAF) that adapts to the frequency ranges of the insects. A CNN (efficient-net) and a transformer (Passt) backbone were added and evaluated. The final models allow researchers of Naturalis to monitor insect populations safely without relying on the current invasive methods.

The customer

Naturalis Biodiversity Center (NBC) is the Netherlands’ national museum of natural history and a leading biodiversity research institute, headquartered in Leiden. With a collection of more than 42 million specimens, Naturalis is one of the foremost natural history institutions in the world. Its research mission centres on understanding, documenting, and preserving global biodiversity, with a strong emphasis on digital tools that make biodiversity data openly accessible to scientists and the public. Our key client contact was Freek de Bruijn. Our team always met biweekly with him at naturalis to discuss the progress of our past sprint and goals for the next one. The communication was always very swift and straight forward resulting in clearly communicated goals. Our contact was always available to give input and help with our challenges we encountered.

The team

We had a Scrum master and a project owner. Work was divided equally. Team members were assigned to tasks alone, but could tag other members for help if necessary. In the beginning the biggest challenge was the knowledge gap in our team, because we had never worked with audio classification before. We solved it by reading a lot of papers about the specific problem and documented our findings. Another challenge was the limited time we had due to long training sessions of the models. We were able to minimize the training time by significantly optimizing the dataloader by utilizing precomputation of audio windows and caching. We are proud that even though it was a tough project, we were able to provide decent results for the client and that we never gave up even when the project was in a bad state.

The technologies

  • Languages: Python, Shell scripting
  • Important frameworks and libraries: pytorch, torchvision, torchaudio
  • Tooling: Ruff, pytest
  • HPC: Snellius (SURF), Slurm