Underwater Acoustic Target Classification Using Scattering Transform with Small Sample Size

Underwater acoustic target classification using passive sonar sensors is a challenging task, primarily due to several factors such as a scarcity of available training data, weak signal power, and the intricate nature of the underwater environment. In the classification pipeline, feature extraction plays a crucial role. However, classical methods have limitations in performance, and deep learning (DL) methods face difficulties in training with a limited number of samples in the practice of this task. To overcome this challenge, we propose a novel approach that leverages the scattering transform for feature extraction. We employ 1-D and 2-D scattering transforms to process raw time-domain signals and time-frequency spectra, respectively. The scattering transform has similar properties to DL methods, but it does not require parameter learning since it is constructed using a multilayer wavelet transform and a predefined nonlinearity. This makes it advantageous, particularly when the training samples are scarce. We use support vector machine as classifier. To evaluate our proposed architecture, we conduct two classification tasks using the ShipsEar database and the DeepShip dataset. The first task involves classifying four highly distinctive types of vessels, referred to as coarse class classification. The second task is to classify three vessels of the same type, referred to as fine class classification. For the classification tasks based on ShipsEar with only 14 training samples per class, our proposed method achieved a classification accuracy of 98.89% for the coarse class using 2-D scattering transform and 92.96% for the fine class with 1-D scattering transform. These figures represent improvements of at least 3.06% and 0.27%, respectively, compared to the best competing method cited in this article. These experimental results demonstrate that, even with a small training set, our proposed architecture achieved superior classification performance.