In this paper, we present the first machine learning package developed specifically for fish calls identification within a specific range (0–500Hz) that encompasses four Caribbean grouper species: red hind (E. guttatus), Nassau (E. striatus), yellowfin (M. venenosa), and black (M. bonaci). Because of their ubiquity in the soundscape of the grouper’s habitat, squirrelfish (Holocentrus spp.) sounds along with vessel noise are also detected. In addition the model is also able to separate grouper species call types. This package called FADAR, the Fish Acoustic Detection Algorithm Research is a standalone user-friendly application developed in Matlab™. The concept of FADAR is the product of the evaluation of various deep learning architectures that have been presented in a series of published articles. FADAR is composed of a main algorithm that can detect all species calls including their call types. The architecture of this model is based on an ensemble approach where a bank of five CNNs with randomly assigned hyperparameters are used to form an ensemble of classifiers. The outputs of all five CNNs are combined by a fusion process for decision making. At the species level, the output of the multimodel is thus used to classify the calls in terms of their types. This is done by species specific deep learning models that have been thoroughly evaluated in the literature on the species concerned here, including transfer learning for red hind and yellowfin groupers and custom designed CNN for Nassau grouper, which has a greater number of known call types than the other species. FADAR was manually trained on a diversity of data that span various regions of the Caribbean Sea and also two recorder brands, hydrophone sensitivities, calibrations and sampling rates, including a mobile platform. This strategy has conferred FADAR substantive robustness to a diversity of noise level and sources that can be found in the grouper calls frequency band such as vessels and marine mammals. Performance metrics based on sensitivity (recall) and specificity showed the same performance level for both balanced and unbalanced datasets and at locations not used in the training set.
