Feature Weighting for Improved Classification of Anuran Calls

Abstract

Automatic bioacoustics monitoring has a great potential to assess the ecosystem health. However, such bioacoustics systems are not highly accurate because the classification of data involves a large number of species. In this paper, we have considered the related problem which involves classification of frog and toad species from their sounds. A publicly available large dataset is used for this purpose where performance is evaluated with leave-one-out cross-validation on the k-NN classifier. The dataset was prepared by extracting Mel-frequency cepstral coefficients (MFCCs)features from the recorded anurans calls, and it comprises the classification of anurans at family, genus and species levels. This paper presents the application of feature weighting to improve the classification of anurans calls. It is a continuous search problem where weights are assigned to features with respect to their contribution in classification. These weights are searched with the Ant Lion optimization along with the best parametric values of the k-NN classifier. The outcomes of experiments show that the proposed approach has successfully enhanced the classification accuracy at family, genus and species levels. The maximum classification accuracies of 95.01%, 88.38%,and 88.08% are achieved at family, genus and species levels respectively which has outperformed the feature selection approach as well as existing works.