Detection and localization of Goliath grouper using their low-frequency pulse sounds.

The goal of this paper is to implement and deploy an automated detector and localization model to locate underwater marine organisms using their low-frequency pulse sounds. This model is based on time difference of arrival (TDOA) and uses a two-stage approach, first, to identify the sound and, second, to localize it. In the first stage, an adaptive matched filter (MF) is designed and implemented to detect and determine the timing of the sound pulses recorded by the hydrophones. The adaptive MF measures the signal and noise levels to determine an adaptive threshold for the pulse detection. In the second stage, the detected sound pulses are fed to a TDOA localization algorithm to compute the locations of the sound source. Despite the uncertainties stemming from various factors that might cause errors in position estimates, it is shown that the errors in source locations are within the dimensions of the array. Further, our method was applied to the localization of Goliath grouper pulse-like calls from a six-hydrophone array. It was revealed that the intrinsic error of the model was about 2 m for an array spanned over 50 m. This method can be used to automatically process large amount of acoustic data and provide a precise description of small scale movements of marine organisms that produce low-frequency sound pulses.

A multimodel deep learning algorithm to detect North Atlantic right whale up-calls.

We present a new method of detecting North Atlantic Right Whale (NARW) upcalls using a Multimodel Deep Learning (MMDL) algorithm. A MMDL detector is a classifier that embodies Convolutional Neural Networks (CNNs) and Stacked Auto Encoders (SAEs) and a fusion classifier to evaluate their output for a final decision. The MMDL detector aims for diversity in the choice of the classifier so that its architecture is learned to fit the data. Spectrograms and scalograms of signals from passive acoustic sensors are used to train the MMDL detector. Guided by previous applications, we trained CNNs with spectrograms and SAEs with scalograms. Outputs from individual models were evaluated by the fusion classifier. The results obtained from the MMDL algorithm were compared to those obtained from conventional machine learning algorithms trained with handcrafted features. It showed the superiority of the MMDL algorithm in terms of the upcall detection rate, non-upcall detection rate, and false alarm rate. The autonomy of the MMDL detector has immediate application to the effective monitoring and protection of one of the most endangered species in the world where accurate call detection of a low-density species is critical, especially in environments of high acoustic-masking.

Transfer learning for efficient classification of grouper sound.

A transfer learning approach is proposed to classify grouper species by their courtship-associated sounds produced during spawning aggregations. Vessel sounds are also included in order to potentially identify human interaction with spawning fish. Grouper sounds recorded during spawning aggregations were first converted to time-frequency representations. Two types of time frequency representations were used in this study: spectrograms and scalograms. These were converted to images, and then fed to pretrained deep neural network models: VGG16, VGG19, Google Net, and MobileNet. The experimental results revealed that transfer learning significantly outperformed the manually identified features approach for grouper sound classification. In addition, both time-frequency representations produced almost identical results in terms of classification accuracy.

Classification of red hind grouper call types using random ensemble of stacked autoencoders.

In this paper, a method is introduced for the classification of call types of red hind grouper, an important fishery resource in the Caribbean that produces sounds associated with reproductive behaviors during yearly spawning aggregations. For the undertaken task, two distinct call types of red hind are analyzed. An ensemble of stacked autoencoders (SAEs) is then designed by randomly selecting the hyperparameters of SAEs in the network. These hyperparameters include a number of hidden layers in each SAE and a number of nodes in each hidden layer. Spectrograms of red hind calls are used to train this randomly generated ensemble of SAEs one at a time. Once all individual SAEs are trained, this ensemble is used as a whole to classify call types of red hind. More specifically, the outputs of individual SAEs are combined with a fusion mechanism to produce a final decision on the call type of the input red hind sound. Experimental results show that the innovative approach produces superior results in comparison with those obtained by non-ensemble methods. The algorithm reliably classified red hind call types with over 90% accuracy and successfully detected some calls missed by human observers.