1. In recent years, seismic sensors, traditionally used in geophysical studies, have been utilized to record seismic waves generated by wildlife locomotion, providing new ways to monitor wildlife non-invasively and continuously. Given the novelty of this approach, numerous research questions, unexplored potentials, and methodological challenges remain. 2. In this study, we investigate the seismic signal properties of African savanna species during locomotion and employ machine learning techniques to classify wildlife based on these footfall signals. We utilize the SeisSavanna dataset, which contains over 70,000 three-component seismograms paired with labelled images from co-located camera traps. To create a graphical overview of the entire seismic dataset, we combine a scattering transform with uniform manifold approximation and projection (UMAP). While the different wildlife categories display different footfall patterns, local geological conditions known as site effects significantly alter the frequency content of those signals. To address the issue of the site effect, we trained machine learning models on data recorded on various sites. 3. For a multi-class classification task involving signals from elephants, giraffes, hyenas, and zebras, the models achieved a balanced accuracy of 87% at a maximum animal-sensor distance of 50 m. The accuracy decreases to 77% when the maximum distance is extended to 150 m due to decreasing signal and label quality. We demonstrate that the models can generalize to new seismic stations if similar site conditions are present in the training data. 4. Our results indicate the potential for using seismic signals in wildlife monitoring and conservation, complementing other existing passive monitoring sensors such as camera traps or acoustic loggers with new observables about silent species. However, further methodological advancements and larger datasets are essential for this approach to become a reliable tool in wildlife monitoring.
