Rumble in the jungle: Convolutional neural networks demonstrate accurate footfall identification of terrestrial mammals

Accurate and resource‐efficient wildlife survey methodologies are crucial to informing biodiversity, wildlife management and conservation planning. Large mammals are particularly challenging to survey due to their large spatial ranges and often‐elusive behaviour. Recent exploration of new approaches has led to seismology being used primarily in elephant behavioural studies; however, little had been done to investigate whether this methodology could provide accurate identification of multiple species and of those with lower body masses. We further developed this seismological approach by utilising convolutional neural networks (CNN) to individually identify four mammal species from recordings of their footfall. To facilitate this work, we built a prototype footfall trap — a seismic node customised for recording footfall — which is both inexpensive and easily replicable. Footfall events were extracted using a fully automated workflow and used to train four CNN models. We extracted 10,965 footfall events from 9 days of recordings in a captive environment and observed vibrations detected within a 3m radius. The best‐performing CNN model achieved a species identification accuracy of 92% with F1 scores for each species between 86 and 97. Disregarding low confidence predictions (<0.8) increased accuracy to 99% but resulted in 27% data loss. The prototype footfall trap costs approximately 220 USD; this is substantially cheaper than seismometers used in similar studies and could be further reduced. Solution. Our findings demonstrate that footfall traps have the potential to provide accurate species identification and could be fully automated once the necessary training data has been established. This methodology leverages some of the benefits of passive acoustic monitoring but for silent species, while overcoming some camera trap limitations in being omnidirectional and unrestricted by surrounding vegetation. Further research is required to understand effectiveness with a greater range of species, particularly taxonomically similar species, and how well the models generalise when facing other environmental conditions.