Marine Mammals Sound Classification Using a Lightweight CNN for Edge Computing Devices

This paper presents a lightweight Convolutional Neural Network (CNN) for marine mammals sound classification. Seven classes are considered in this work, namely humpback whale, killer whale, pilot whale, sperm whale, spinner dolphin, striped dolphin, and white whale. The sound signals are transformed from the .wav file to time-frequency images through Continuous Wavelet Transform (CWT). The proposed CNN consumes time-frequency images as input and consumes only a 0.8 MB size model benchmarked against several well-known pre-trained models, including DenseNet, Efficient-Net, InceptionNet, MobileNet, NASNet, ResNet, VGGNet, and Xception. Experimental results demonstrate that the proposed lightweight CNN significantly outperformed pre-trained models, able to achieve a validation accuracy of 80.9% in a ten-fold cross-validation framework. The best performance among the pre-trained models is Mo-bileNetV3Small, which achieved only 78.91% validation accuracy in a ten-fold cross-validation framework with a considerably larger model size of 4 MB. Furthermore, the proposed model was successfully deployed on the Raspberry Pi IOT module and able to run with 30.134 milliseconds per inference. The lightweight CNN’s high accuracy and low latency make it an ideal candidate for applications in real-time audio event detection in underwater with resource-constrained devices.