Wildlife Classification using Acoustic Features and Deep Learning Approach

Wildlife conservation is vital for maintaining the balance between ecosystems and preserving biodiversity securing the survival of diverse species and their habitats. It mitigates the climate change impact and safeguard the natural and ecological services important to human life. This paper presents a pre-trained model approach and a deep learning approach for classification. In the pre-trained model approach, feature extraction is based on Mel-frequency cepstral coefficients (MFCCs) and Linear Predictive Coding (LPC) which are trained and tested on pre-trained models namely MobileNetV2, EfficientNet, ResNet50, ResNet101. In the deep learning approach for classification, features are extracted based on MFCC using an FFT window of length 2048, 128 Mel bands and hop length 512. A custom CNN is used to further classify the MFCC’s into their respective classes. Exploratory data analysis (EDA) and preprocessing techniques are used to delve into the dataset’s composition and structure, addressing challenges such as class imbalances and background noise. An accuracy of 93% is achieved for classification of the species on the Rainforest Connection Species Audio Detection dataset(RFCx dataset).