Schooling machine learning to identify fish
The National Oceanic and Atmospheric Administration wants to leverage machine learning to improve its ability to verify marine species.
NOAA’s Fisheries Sampling Branch is responsible for monitoring and observing fishing in the Northeast and mid-Atlantic. FSB collects, processes and manages data and biological samples on fishing operations, catch, including bycatch and discard information to provide accurate near-real-time data for fisheries management and evaluating the status of marine life
Over the last 25 years, FSB has worked to make this identification easier for its observers -- who started out taking pictures of samples of frozen fish with film cameras, noting the location and focusing on the unique characteristics that allow differentiation of each fish species, such as whole body, gills or fins. The images were submitted within 48 hours and verified by other observers.
Currently, images -- which still vary greatly in quality due to different cameras, photographic conditions and characteristics of the species -- are upload to the Fish House, a web-based user interface for a Oracle relational database that is used for species verification.
According to a recent request for quotations, FSB wants a contractor to develop open-source machine learning software trained on an existing library of previously identified photos. It must not only confidently identify 45 common fish species but also improve as new images and verification data become available and incorporate the improvements into its existing species verification process.
Assuming some or most species can be reliably identified, this solution might allow partial or full automation for these identifications, reducing the time required by FSB staff for image verification review.
Responses are due Aug. 14 -- the Friday of "shark week."
Connect with the GCN staff on Twitter @GCNtech.