Publication

Abstract : Classification of sounds from marine mammals is an essential challenge for successful ecological monitoring and conservation. However, the complexity in these sounds makes them more challenging to classify. This paper presents a two-phase approach that integrates traditional machine learning techniques with pretrained deep learning models. In the first phase, pretrained networks-namely MobileNetV2, EfficientNetV2B0, ResNet18, and InceptionV3 were used to extract features from spectrograms, including Mel-frequency cepstral coefficients (MFCC), fast Fourier transform (FFT), and constant-Q transform (CQT). These extracted features were then classified using various machine learning algorithms to accurately distinguish 23 distinct sound classes of marine mammals. Experimental results revealed that the light weight ResNet18 model provides a comparable performance with respect to heavier models such as MobileNetV2, InceptionV3 and EfficientNetV2B0. Building on these findings, the second phase focuses on designing a modified ResNet18 architecture for the classification of marine mammal sounds. Exhaustive hyperparameter tuning was conducted, leading to improved performance. The modified ResNet18 achieved a notable accuracy of 98.32%. These experimental findings highlight the importance of customizable deep learning architectures and thorough hyperparameter tuning to enhance the accuracy of marine mammal sound classification.