Marine Mammals Sound Classification using Extreme Learning Machine with Kepler Optimization Algorithm

This paper presents a new collaborative technique for classifying marine mammals sound, which includes both texture and audio features. A novel feature descriptor called 1DLOTBP is proposed to effectively extracts the features from the audio spectrogram. Along with the proposed 1D-LOTBP other feature descriptors namely LBP, 1D-LBP, ELBP are combined to extract information from audio spectrogram images. To further improve the performance of the texture descriptor the audio features like MFCC, CQT, ZCR and STE added. To validate the proposed collaborative technique, the input feature values are imported to ELM classifier. The ELM is a single layer feed forward neural network, which randomly generates weights and bias in the network. In order to overcome the above drawback of randomness the ELM classifiers are combined with optimization technique called Kepler optimization algorithm (KOA). KOA improves the exploration and exploitation search space rate because the candidate solutions show distinct conditions from the Sun at different times. The proposed technique is validated using the whale sound database. In first stage, the single features are imported to know the individual classification accuracy, then the features are collaborated to improve the accuracy further. The proposed KOA-ELM classifier achieves higher classification accuracy of $\mathbf{9 9 \%}$ using the collaborative feature descriptor.