WildTechAlert: Deep Learning Models for Real-Time Detection of Elephant Presence Using Bioacoustics in an Early Warning System to Support Human-Elephant Coexistence

Human-elephant conflict (HEC) in Malaysia occurs in areas where humans and Asian elephants (Elephas maximus) share landscapes, especially at forest edges and agricultural lands. Due to increased anthropogenic activities, such as logging, agriculture, and infrastructure developments, a large portion of natural habitat has been converted and fragmented, resulting in increased interactions between people and wild elephants. This presents a unique challenge of conserving elephants, safeguarding human livelihoods and ensuring habitat connectivity for endangered wildlife. The ability to detect and predict the presence of elephants in human-dominated landscapes can help increase safety for communities and provide insights for decision-makers to better manage conflict situations and promote a harmonious human-elephant coexistence. In the past decade, technologies used, such as camera traps, drone imaging, and GPS tracking, have allowed us to understand activity patterns, habitat preferences, and behaviour of elephants, but they often have limitations in terms of cost, coverage effort, and manpower needed. The increasing application of artificial intelligence (AI) in conservation offers an opportunity to develop real-time solutions to detect elephant presence in areas with a high risk of human-elephant encounters and HEC. In this study, we present a prototype of an early warning system, WildTechAlert, which combines bioacoustics and AI technology. Bioacoustics is the study of biological sounds; and the prototype device, being able to record sounds travelling from all angles, may potentially have higher detection rates from its surroundings (omnidirectional coverage) than sensors with fixed angles, such as camera traps. A sound stored in the raw waveform format can be converted into a spectrogram, which is a visual representation of the sound signature in the frequency, time, and amplitude domains. The spectrograms of elephant sounds were used as images for training a deep learning model using convolutional neural networks. WildTechAlert includes a device equipped with bioacoustic sensors which connect to a cloud-based deep learning algorithm, a user-friendly web interface, and mobile notification alert functions. The alerts serve as a tool to increase the safety of communities and plantation workers and promote preparedness for mitigation actions. We trained binary and multiclass classifiers to detect elephant sounds and achieved up to a 94% accuracy in detecting elephant sounds with the binary classifier. The performance of this system has yet to be tested in field conditions but shows potential in HEC management and to promote coexistence between humans and elephants.