Enhanced Frequency Domain Analysis for Detecting Wild Elephants in Asia using Acoustics

Human-elephant conflict in Asia and Africa calls for an early warning system to reduce risks and harm for both elephants and humans. Acoustic based warning systems offer a promising solution due to their non-invasive and cost-effective nature. In this paper, we propose a novel approach for detecting wild elephants using acoustic signals, targeting the Asian elephant population in Sri Lanka. The proposed method introduces a unique data preprocessing technique, followed by feature extraction using a deep convolutional neural network followed by fully connected layers for classification. Spectro-grams are used as input data, and transfer learning is employed with YAMNet model layers. Additionally, we have developed a hardware system capable of capturing infra sound signals, although a detailed description of the system is beyond the scope of this paper as it is crucial for detecting elephant activity. The proposed method is evaluated on a large data set recorded under natural field conditions in Sri Lanka, and it demonstrates 97.77% accuracy in detecting elephants and robustness to noise sources. Proposed approach has the potential to develop into a non-invasive early warning system for elephant detection in the wild, contributing to the mitigation of human-elephant conflict and wildlife preservation.