Did U hear that? Working with mixed behaviours when classifying animal behaviour from acceleration data using a U-Net
Abstract
Current approaches to detect roaring behaviour of lions ( Panthera leo ) within acceleration data are scarce and limited to males that are stationary. We propose a segmentation approach using a Fully Convolutional Neural Network (U-Net) that can also work with data from females and roars that happen while the lion is walking. We equipped seven lions with a GPS/accelerometer and an audio logger. We merged audio and acceleration data to identify roaring signals in the acceleration data. We then trained a U-Net to differentiate between “roaring” and “no roaring” data. As training data we used segments containing only “no roaring” data and segments with a mix of “roaring” and “no roaring”. Finally, we used our classifier on the complete data from five lions to test how well the classifier would perform when roaring events only make up a small portion of the data. Our model shows a precision for “roaring” data points of 0 . 94 ± 0 . 01 (mean ± 95 % confidence interval) and a recall of 0 . 73 ± 0 . 01 over all individuals. Based on this, we could correctly identify between 90% and 96% of all roaring bouts with a precision between 87% and 100%. We showed that our approach can detect lion roaring behaviour of males and females in a stationary and walking context reliably. This can broaden the context in which lion communication can be studied and thus provide less biased data.
Related
- Processing of acceleration and dive data on-board satellite relay tags to investigate diving and foraging behaviour in free-ranging marine predators
- A standardised framework for classifying estimates of reproductive isolation
- Shock Detectives: classifying MMS observations of Earth’s magnetosheath via citizen science
- A deep feature-based approach for classifying flue gas pollutants and hazard identification using pyramid-dilated temporal convolution networks
- Ontogeny of settlement behaviours in response to Grammatophora marina diatom biofilms in the marine polychaete, Platynereis dumerilii
BibTeX
@article{rast2026hear,
title = {{Did U hear that? Working with mixed behaviours when classifying animal behaviour from acceleration data using a U-Net}},
author = {Wanja Rast and Theresa Götz and Claudine Cloete and Anne Berger and Simon Chamaillé-Jammes and Miha Krofel and Rubén Portas and Ortwin Aschenborn and Jörg Melzheimer},
journal = {Ecological Informatics},
year = {2026},
month = nov,
doi = {10.1016/j.ecoinf.2026.103761},
url = {https://hal.science/hal-05853045v1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Did U hear that? Working with mixed behaviours when classifying animal behaviour from acceleration data using a U-Net
AU - Wanja Rast
AU - Theresa Götz
AU - Claudine Cloete
AU - Anne Berger
AU - Simon Chamaillé-Jammes
AU - Miha Krofel
AU - Rubén Portas
AU - Ortwin Aschenborn
AU - Jörg Melzheimer
PY - 2026
DA - 2026/11/05
JO - Ecological Informatics
DO - 10.1016/j.ecoinf.2026.103761
UR - https://hal.science/hal-05853045v1
AB - Current approaches to detect roaring behaviour of lions ( Panthera leo ) within acceleration data are scarce and limited to males that are stationary. We propose a segmentation approach using a Fully Convolutional Neural Network (U-Net) that can also work with data from females and roars that happen while the lion is walking. We equipped seven lions with a GPS/accelerometer and an audio logger. We merged audio and acceleration data to identify roaring signals in the acceleration data. We then trained a U-Net to differentiate between “roaring” and “no roaring” data. As training data we used segments containing only “no roaring” data and segments with a mix of “roaring” and “no roaring”. Finally, we used our classifier on the complete data from five lions to test how well the classifier would perform when roaring events only make up a small portion of the data. Our model shows a precision for “roaring” data points of 0 . 94 ± 0 . 01 (mean ± 95 % confidence interval) and a recall of 0 . 73 ± 0 . 01 over all individuals. Based on this, we could correctly identify between 90% and 96% of all roaring bouts with a precision between 87% and 100%. We showed that our approach can detect lion roaring behaviour of males and females in a stationary and walking context reliably. This can broaden the context in which lion communication can be studied and thus provide less biased data.
SN - 1574-9541
ER -