An Efficient Marine Sound Classification Based on Fourier Spectrogram of Compressive Measurements and One Dimensional CNN with Optimal Layers

The underwater environment is a complex acoustic habitat, rich in biotic and abiotic sounds. Detecting and classifying diverse ocean sounds is essential for understanding and protecting marine ecosystems, monitoring the health of the ocean, and managing human activities that impact marine life. However, it remains challenging because of their time-varying temporal-spectral characteristics. Most deep learning-based marine sound classification methods were presented with a better model size and parameters using the uncompressed sound and spectrogram of the uncompressed sound signal. In this paper, we present computationally efficient marine sound classification based on the Fourier spectrogram of compressive measurements and a one-dimensional convolutional neural network (1D CNN) with an optimal number of layers. In this study, we consider three sensing matrices to achieve data reduction that include the random Bernoulli sensing matrix (RBeSM), the random sparse binary sensing matrix (RSBSM), and the random sparse fixed binary sensing matrix (RSFBSM). Among the sensing matrices, the RBeSM-based 1D CNN method had the highest accuracy of 91.45%, specificity of 81.01% and sensitivity of 99.98%. When comparing the performance of the CNN model with 2-layer, 3-layer and 4-layer architectures with the ReLU function, the 4-layer CNN RBeSM model achieved the highest sensitivity of 99.8% and an accuracy of 91.45% as compared with other sensing matrices. The evaluation results demonstrate that the compressive measurements-based method can reduce the computational load, leading to a reduction in latency and energy consumption.