ODF Sweden has founded a new platform/tool for the Koster Project where you upload images and analyze marine ecological research from the Koster Seafloor Laboratory. The output is not just about identifying one specific species, but also a platform/tool where you can have access to others movies and upload your own of the Lophelia Pertusa corals. Both people who are familiar with marine fauna and beginners can contribute.
How to use the application:
The app has two modes, one specifically for looking through past footage used for training the model, and another which allows users to upload their own footage by dragging and dropping files or clicking on “Browse files”. Users can switch between these modes by clicking on the “Custom File Upload” tick box located in the sidebar.
Once the footage has been selected, users can change the confidence and overlap thresholds to see how these affect the model output. The confidence threshold determines which bounding boxes are kept (filtering out those with lower confidence) and the overlap threshold determines which boxes should be combined into a single bounding box. This allows the users to fine-tune the process of outputting the correct number of bounding boxes in the correct position. When viewing previously uploaded footage, users can also select from a range of movies and then from a range of frames to see the various annotated frames.
The app will be released soon and available here.
The aim of this app is to make the project’s outputs more accessible and flexible. The non-technical audience can upload and their own footage and get predictions directly from the model. Users can also tweak important information that affects the model output, such as the confidence threshold and overlap thresholds of various bounding boxes.
Screenshot from the web app.
The key takeaways, ODF Sweden aims to continue adding footages for the seafloor development as more protected areas get added. However, the usage of machine learning is still not perfect and they still face many challenges in future iterations.
Repositories associated with the KSO:
Github repository for the machine learning model
Github repository for the system data flow
Previous Research
The application was evolved from previous research by Victor Anton, Jannes Germinshuys, and Matthias Obst. Their research in the Koster Seafloor Observatory were developed from a customized citizen science website that used a machine-learning algorithm made from classifications by scientists, allowing researchers to track biological objects in new footage and facilitating the analysis of the subsea. The systems were possible by recordings from remotely operated vehicles (ROV) that monitored the Kosterhavets National Park and analyzed its highly diverse and unique marine reserves.
The results showed that cold-water corals have strongly declined during the last 15 years due to changing water temperature and fishing. There seems to be no recovery of the coral stocks since the establishment of the national park in 2009. There are other factors essential for creating a recovery of the ecological species for example external climate pressures, changes in water quality, and oceanographic connectivity possibly have a large impact on the coral population.
Screenshot from workflow 1 of Koster Seafloor Observatory on Zooniverse.
Ocean Data Factory is one of the founders of The Koster Project and is strongly impressed by its efficency. It was also supported by the Swedish Agency for Marine and Water Management, the Swedish Research Council, the NEIC program DeepDive and the Horizon 2020 project ENVRIpls.
The challenge for the upcoming web app is to expand the model, include more key species and to improve the image quality of older footages. Hopefully, this new application will add more areas and more spices to the open-source database.
Read the full article about the previous work in the Koster project here.
Read more about the new web app here.
1 thought on “Marine Species Recognition: a Web app”