Analyzing Frog Vocalizations Using MFCCs and Machine Learning for Species, Family, and Genus Identification

The dataset analyzed and described here involves audio recordings Mel-Frequency Cepstral Coefficients, with taxonomic labeling at family, genus, and species levels from frogs. The main focus of the proposed work is to perform frog classification using machine learning in an attempt to rectify the class imbalances intrinsic in the data set. To achieve this objective, preprocessing was done: categorical encoding and most informative feature selection through statistical techniques. Class weighting was applied for the class imbalance problem in balanced learning. In terms of accuracy, F1 score, and recall, several models were trained in machine learning: Logistic Regression, Random Forest, Gradient Boosting, and XGBoost. Feature selection improved the model performance as it reduced the dimensionality and picked the top 10 features relevant to classification. The results of this study clearly show that ensemble-based models have achieved better performance with the highest accuracy at $99.68 \%$ and an F1 score at 0.99, even with the presence of an imbalanced class. This approach provides a solid framework for using MFCCs in bioacoustic research and brings to light the prospect of machine learning for the analysis of ecological data.