Large-scale automatic acoustic monitoring of African forest elephants' calls in the terrestrial acoustic recordings

Abstract

African forest elephants live in the rain forests of western and central Africa. The dense habitat prevents them from communicating visually within the family group. Automatic detection of African Forest Elephants' calls intercepts signals in their communication channel and enables fast processing of large scale acoustic data. In this work, first, an automatic detection system targeting at African forest elephants' rumble calls is proposed. De-noising pre-processing, design of acoustic feature vectors, and choice of classifiers are discussed respectively. Second, the detector's performance is evaluated by the cross-validation of a 432-h of acoustic recording from eight locations in Gabon, Africa. It shows that the detector achieve 79.19% true positive rate when the false positive number is the low 5.70 per hour. The F1-score (geometric mean of precision and recall) is around 0.77 when relatively high score threshold (over 0.8) is selected. Finally, a case study demonstrates the results of applying our automatic detection system to a large-scale data set, which amounts to 420 days of acoustic recording over 3 years from the Ivindo National Park, Gabon. Visualization of the call activities reveal the seasonal and daily patterns as well as the temporal variation over the 3 years.