EcoTaxa: a tool to support the taxonomic classification of large datasets through supervised machine learning - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

EcoTaxa: a tool to support the taxonomic classification of large datasets through supervised machine learning

Résumé

Images are increasingly used as a means to collect data in all fields of science, and Ecology is no exception. In the underwater realm, where direct observation of the organisms in their environment is difficult for humans, automated cameras provide invaluable insights. This partly explains the flourish of camera-based instruments specialised for taking pictures of plankton. Because most of them image a controlled volume in a systematic manner, they are coined "quantitative imaging" instruments; they allow computing concentrations and making replicable morphological measurements on the many thousands of images they collect. This also opens the avenue for the automation of their classification. EcoTaxa was designed as a platform to upload images, together with rich metadata, and sort them taxonomically in an efficient way. This efficiency is partly provided by machine learning: users can train models based on previous identifications in the database to suggest labels for newly uploaded images. By combining deep-learning feature extractors, a fast-to-train classifier, and enough flexibility to train models customised to the task at hand, EcoTaxa achieves classification performance similar to that of state of the art deep-learning networks while being usable in a matter of minutes by taxonomists with no computer science knowledge. The efficacy is also provided by the web-based graphical user interface: several users can collaborate on the classification of a dataset and each can rapidly review and classify hundreds of images at a time. As a result, trained operators routinely sort 5,000 to 10,000 per working day, within ~100 taxonomic groups. In the application as a whole, over 200 million images have been uploaded and over 90 million have been sorted by human operators, in its 6 years of existence. We will review the principle, functioning and potential for generalisation of the approach implemented in EcoTaxa.
Fichier non déposé

Dates et versions

hal-04026447 , version 1 (13-03-2023)

Identifiants

  • HAL Id : hal-04026447 , version 1

Citer

Jean-Olivier Irisson, Laurent Salinas, Sebastien Colin, Team Complex, Marc Picheral. EcoTaxa: a tool to support the taxonomic classification of large datasets through supervised machine learning. SFEcologie 2022, 2022, Metz, France. ⟨hal-04026447⟩
210 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More