Seismic Footsteps: Harnessing Machine Learning to Decode Wildlife in the African Savanna

Seismic sensors, traditionally used in geophysical studies, are emerging as non-invasive tools for continuous wildlife monitoring by capturing seismic waves generated by animal locomotion. This novel approach opens new possibilities but also presents methodological challenges. In this study, we analyze seismic signals from African savanna species during locomotion and apply machine learning to classify species based on footfall signals. Utilizing the SeisSavanna dataset, which includes over 70,000 labeled seismograms paired with camera trap images, we identify distinct species-specific footfall patterns. Our analysis reveals that local site effects significantly influence signal frequency content. To address this, we trained machine learning models on data from multiple locations, achieving a balanced accuracy of 87% for elephants, giraffes, hyenas, and zebras at distances up to 50 meters, decreasing to 77% at 150 meters due to weaker signals and lower label quality. Importantly, the models generalize well to new stations if similar site conditions are represented in the training data. These findings highlight the potential of seismic monitoring to complement tools like camera traps and acoustic loggers, offering unique insights into wildlife behavior and expanding monitoring capabilities to silent species. To fully realize this potential, further methodological advances and larger datasets are necessary to establish seismic sensors as a robust tool for wildlife conservation.