Roaring over long distances is a key behaviour of lions. They communicate within prides as well as with other animals using distinct sequences of moans and grunts. Scientists from the GAIA Initiative have now published a machine learning approach in the journal Ecological Informatics that improves how roaring behaviour can be studied. The algorithm can reliably detect long-distance roaring based solely on acceleration data (ACC) that is recorded by collars – without a microphone and without energy- and storage-intensive audio files. For the first time, such an algorithm works reliably with both male and female lions, and even with mixed signals when lions are walking while roaring.
Lions live in highly social groups (prides) with close bonds and a complex communication system comprising visual, chemical, physical and acoustic signals. They use these signals to coordinate their movements and avoid conflicts. The only way for lions to communicate over long distances is by roaring. As the members of a pride are often scattered over large distances, roaring plays a central role in communication within the pride. Although the acoustic details of the roaring sounds have already been thoroughly documented and analysed, the spatial behaviour associated with roaring and the functions of roaring within species-specific communication have only been studied to a limited extent. “It is known that male lions do not roar to impress female lions, and it is suspected – though as yet without scientific evidence – that lions roar to maintain their territory,” says Dr Ortwin Aschenborn, wildlife veterinarian and scientists in the GAIA Initiative at the Leibniz Institute for Zoo and Wildlife Research (Leibniz-IZW). “Long-distance communication in female lions remains completely unexplored.”
Traditional audio recording techniques are not practical for this type of research: handheld microphones can only capture individual lion vocalisations sporadically from a distance. Audio recording devices attached to the animals, such as “audio loggers” combined with collars, require a great deal of energy and storage space – and record large amounts of data that are not needed for answering these research questions. To analyse the roar in combination with spatial movement patterns and thus social behaviour, it is necessary to record over long periods when and where a lion (or lioness) roars. This information can be derived from the analysis of so-called acceleration data (ACC) that is recorded by collars. ACC data records the even the slightest movements in three dimensions in rapid succession and over long periods, making it possible to identify a wide range of behaviours. Artificial intelligence is an indispensable aid in this process: machine learning algorithms can be trained to identify specific patterns in the acceleration data and assign them to a particular behaviour. Movement patterns such as running, flying or swimming can already be detected very reliably for many animal species and studied scientifically, for example in conjunction with the animals’ GPS position recorded at the same time. However, classification of vocalisations from acceleration data is still rare, as it presents significant methodological challenges.
GAIA’s “fully convolutional neural network” can identify a lion's roar in ACC data even when it occurs alongside other behaviours such as running
Unlike running or flying, movement patterns associated with vocalisations are linked only to subtle, fine-scale movements that often affect only part of the animal’s body. Although a lion’s roar is very loud (it can be heard from up to eight kilometres away), the resulting ACC signal is relatively weak. Furthermore, the vocalisation can also be mixed with other behavioural patterns, for example when a lion – as regularly happens – moves whilst roaring and the sensors record a mix of different behavioural patterns. “Previous models for classifying lion roars from acceleration data were trained exclusively on male lions that were not moving in any other way,” explains Wanja Rast, wildlife AI specialist in the GAIA Initiative and PhD student at the Leibniz-IZW. “Our new development can do more: we have trained a so-called ‘U-Net’ that can detect the roars of both male and female lions – both whilst they are moving and when they are stationary.”
A “U-Net” is a variant of a “convolutional neural network” (CNN), a machine learning approach primarily used for image and audio files. The data is organised into layers that are processed through convolutions, generating an output value (in the case of the lion study, “roar” or “no roar”) from an input value. The model was trained using reference data from seven collared lions in Etosha National Park, which wore both a GPS collar with accelerometer and an audio logger over a period of several months. In total, the scientists recorded 1,333 roaring events. Acceleration and audio data were synchronised so that the roaring signals could be identified in the ACC data stream. After training, the U-Net was able to perform this classification with an accuracy of 90 to 96 percent based solely on acceleration data – the AI missed only a few of the actual roars (false negatives). Conversely, around 81 percent of the roars flagged were genuine; in just under 20 percent of cases, the AI was incorrect (false positive). Both figures apply equally to male and female lions. “Roars whilst walking were classified slightly less reliably on average; however, subsequent filtering steps were able to improve detection to a level comparable to that of roars without walking”, says Rast.
Collars with ACC sensors can – for specific applications – replace audio loggers on lions and open up new opportunities for research
The GAIA scientists at Leibniz-IZW are convinced that, for certain research questions, it is both feasible and reasonable to rely on AI-assisted classification of acceleration data to investigate animal behaviour in connection with vocalisations. “Unlike audio loggers, which store actual sound recordings, acceleration data can be saved over longer periods using less energy and storage space,” says Dr Jörg Melzheimer, a scientist in the GAIA Initiative at the Leibniz-IZW. However, it is not the actual sound – roaring, grunting or howling – that is recorded, but only the fact of when and (in combination with the GPS signature) where it was produced. “In addition, a well-trained classification model for acceleration data can also be applied to existing datasets, meaning data could be utilised for studies on vocalisation behaviour that were not originally created with this research topic in mind,” says Melzheimer. “At the same time, it must be noted that this method only works if an AI or machine learning algorithm can be successfully trained.” The GAIA Initiative team has succeeded in doing this for lion roars – however, it is possible that attempts with other species may be less successful or fail entirely, as not every vocalisation is associated with characteristic or sufficiently pronounced ACC signals.
Building on this machine learning approach, the GAIA team intends to further investigate lions’ roaring behaviour as a key aspect of intraspecific communication. The scientists also plan to develop a concept for an “acoustic fence” for the boundaries of protected areas, in which strategically placed sensors and loudspeakers in the landscape would tune into lion communication at crucial moments. The aim is to keep lions within these protected areas and reduce contact and conflicts with humans.
Ecological Informatics
Data/statistical analysis
Animals
Did U hear that? Working with mixed behaviours when classifying animal behaviour from acceleration data using a U-Net
20-Apr-2026