With BlueGreen Labs we are change makers, we push science forward without gatekeeping resources or knowledge. We provide what is asked, but also what is needed.
Past spring we were asked to give a workshop on citizen science within the context of data recovery from natural science collections of the Institute of Natural Sciences in Belgium. With BlueGreen Labs we have run three citizen science projects, most prominently Jungle Rhythms, Jungle Weather, and contributed to others. With this workshop we wanted to inform people on how to best use the Zooniverse citizen science platform.
Although we did teach on the strengths and pitfalls of citizen science we also pitched a machine learning approach. It is our opinion that citizen science is valid science, not mere outreach. Therefore, the time of volunteers should not be used gratuitously. All other options to automate a task should be explored first.
In the context of the workshop the scope was to count and classify insects in the entomology collection of the Institute of Natural Sciences (see example image below). However, given advances in all aspects of machine learning this did not look like a job for citizen science anymore. Using a small training set, annotated using an offline instance of CVAT, we trained a small YOLOv8 object (insect) detection model using a couple of hours of work. The model easily attained an accuracy of ~85% with minimal effort using open source tools (a larger model and more training data would increase accuracy further).
At the end of the workshop we demonstrated this approach by showing a small online web application (running the model in the browser). Doing so we made the participants, who previously only considered citizen science to be the solution to their annotation problem, to reconsider machine learning as a valid option.
We provided what was asked, but also what was needed.
