Evaluation of a machine learning approach for underwater target classification with low-frequency activity sonar robust to environment differences

Low-Frequency Active Sonar (LFAS) systems are the method of choice for long-range underwater
surveillance and detection of targets. To extract relevant sonar contacts, the raw LFAS data
undergoes several signal processing steps. These include processes such as beamforming, matched
filtering, detrending of the signal as function of range, and additional analyses to extract candidate
sonar contacts [1]. Each candidate contact is classified as either background/clutter or a potential
target.
Feature-based classification and machine learning or artificial intelligence technology can assist with
(semi)automatic classification of sonar contacts. Such systems can be trained with a dataset of
measured target data [2] [3] or employed as anomaly detection trained only in a given environment
[4]. In both situations, systems heavily rely on the provided recorded example data. Deploying a
trained system in a new underwater environment will likely lead to degraded performance. However,
new LFAS target data is expensive to obtain, environmentally dependent, and scarce. Simulating
sonar data in different environments is a useful approach to generate new data [5]. However, the key
question remains: What is the optimal simulation and training strategy given the environmental
dependency?
In this work, three different approaches of applying trained machine learning systems in a new
underwater environment are studied. Various underwater environments will be simulated and include
clutter (simulated solid sphere made of rock) and targets (simulated air-filled steel spheres). Our
approaches include: (1) a zero-shot/generalization approach that is evaluated in an environment not
present in the training data, (2) a few-shot approach in which the zero-shot classifier is fine-tuned with
a subset of in-situ sonar data, and (3) an environment-specific approach that is only trained on a
subset of in-situ sonar data and evaluated in that same environment. A number of machine learning
systems will be investigated, including a support-vector machine (SVM), a fully-connected neural
network (FCN), and a convolutional neural network (CNN).