| Acoustic surveys for marine fish in coastal waters typically involve identification of species groups. Incorrect classification can limit the usefulness of both distribution and biomass estimates. Fishing catch data can assist in identification, but are rarely spatially comparable to acoustic data and are usually biased by gear type. This thesis describes a technique and a software toolkit, "FASIT" (Fisheries Assessment and Species Identification Toolkit), developed by the author to enable automated identification of Atlantic cod (Gadus morhus), capelin (Mallotus villosus), and redfish (Sebastes spp.) based on high resolution acoustic imaging of fish aggregations. The approach has been to assess and analyze various amplitude, shape and location features of the acoustic returns from shoals and individual fish, then to use these features to develop algorithms which discriminate among species. Fourteen classifiers based on Three-Nearest Neighbour classification and Mahalanobis distance classification have been implemented and tested. The best classifier had an average correct classification rate of 96.8%. The data used for this thesis are fisheries data from a number of Newfoundland bays and the Grand Bank region collected using a 38 KHz digital echo-sounder. |
