The accurate detection and quantification of submerged targets has been recognized as a key challenge in marine exploration, one that traditional census approaches cannot handle efficiently. Here we present a deep learning approach to detect the pattern of a moving fish from the reflections of an active acoustic emitter. To allow for real-time detection, we use a convolutional neural network, which provides the simultaneous labeling of a large buffer of signal samples. This allows to capture the structure of the reflecting signal from the moving target and to separate it from clutter reflections. We evaluate system performance both on synthetic (simulated) data, as well as on real data recorded over 50 sea experiments in a variety of sea conditions. When tested on real signals, the network trained on simulated patterns showed non-trivial detection capabilities, suggesting that transfer learning can be a viable approach in these scenarios, where tagged data is often lacking. However, training the network directly on the real reflections with data augmentation techniques allowed to reach a more favorable precision-recall trade-off, approaching an ideal detection bound. We also evaluate an alternative model based on recurrent neural networks which, despite exhibiting slightly inferior performance, could be applied in scenarios requiring on-line processing of the reflection sequence.

An effective method for detecting the presence of dolphins is by using passive acoustic monitoring (PAM), where pod size indications can be estimated by counting individual whistles. The detection of dolphin whistles is commonly applied on a time-frequency representation, followed by denoising and whistle tracking to evaluate the number of whistles. However, due to harmonics, multipath and time-varying signal-to-noise ratio, a single dolphin whistle may be associated with multiple whistle-traces. Thus, as a first step towards evaluating dolphins' abundance, our goal is to cluster individual whistle traces into unique whistles. Our scheme measures the similarity between each pair of whistle traces, and estimates the likelihood of whistle traces sharing the same cluster. Clustering is formalized as an optimization problem, aims to maximize the stability of clusters. Formalizing the problem as a minimal-cut optimization on a graph provides an effective solution based on spectral decomposition of the graph-Laplacian. Our model of the likelihood sharing cluster provides a physically-meaningful method to calculate the graph's connectivity parameters, thereby leading to a robust blind clustering. Based on numerical simulations and real recordings of dolphin whistles at sea, we demonstrate the applicability of our solution and its advance beyond alternative approaches.

We address the challenge of detecting an arbitrary-shaped underwater acoustic signal. Instead of setting a detection threshold, which due to noise transients may result in a high false alarm rate (FAR), our method classifies each measured sample as either “noise” or “signal.” Utilizing a priori knowledge of only the minimal duration of the signal, the decision is made using loopy belief propagation over a factor graph. Numerical simulations and sea experimental results show that our scheme achieves a favorable tradeoff between the Recall and FAR, and noise robustness, which far exceeds that of benchmark schemes.

We focus on the detection of sporadic low-power acoustic/seismic signals of unknown structure and statistics, such as the detection of sound produced by marine mammals, low-power underground signals, or the discovery of events such as volcano eruptions. In these cases, since the ambient noise may be fast time varying and may include many noise transients, threshold-based detection may lead to a significant false alarm rate. Instead, we propose a detection scheme that avoids the use of a decision threshold. Our method is based on clustering the samples of the observed buffer according to a binary hidden Markov model to discriminate between “noise” and “signal” states. Our detector is a modification of the Baum-Welch algorithm that takes into account the expected continuity of the desired signal and obtains a robust detection using the complex but flexible general Gaussian mixture model. The result is a combination of a constrained expectation-maximization algorithm with the Viterbi algorithm. We evaluate the performance of our scheme in numerical simulations, in a seimic test, and in an ocean experiment. The results are close to the hybrid Cramér-Rao lower bound and show that, at the cost of some additional complexity, our proposed algorithm outperforms common benchmark methods in terms of detection and false alarm rates, and also achieves a better accuracy of the time of detection. To allow reproducibility of the results, we publish our code.