A location fix tells a researcher where a giraffe was. It does not explain what happened there. The animal might have stopped to browse, stood alert or continued chewing food it had already swallowed. Even a detailed movement trace needs interpretation before it can answer a behavioural question.
Stefanie Brandes, Florian Sicks and Anne Berger tested that interpretation with three captive giraffes. They attached acceleration sensors to the animals' heads and watched their behaviour at the same time. The aim was to build a translation between recorded motion and observed activity that might eventually help researchers follow giraffes they could not continually see.
The strongest part of the experiment was a change in the test itself. Instead of asking only whether a computer could recognize more recordings from a familiar giraffe, the team asked it to classify recordings from another animal. Performance declined. That gap describes a central difficulty of turning a promising demonstration into a field method.
The head provides a rich but incomplete signal
Acceleration sensors recorded movement along three perpendicular axes. Their readings also reflected orientation relative to gravity. A head raised toward high leaves produces a different pattern from a head lowered toward the ground; walking adds another kind of motion.
The researchers worked with two females in Canada and one male in Berlin. These giraffes had been trained to accept the equipment, so fitting it did not require sedation. The two females carried both devices being compared, while the male carried one. The sensors differed in recording frequency and how their data were stored or transmitted.
For one device, brief high-frequency bursts were recorded every 20 seconds. Observers used a signal marking the bursts to identify what the giraffe was doing at those moments. The other device recorded readings each second, which the team matched to observed stretches of behaviour.
The classification scheme distinguished feeding at different head heights, standing, walking, rumination and other activities. That detail was biologically useful, but it also made the problem harder. Chewing cud and chewing newly collected food can move the head similarly. A motion sensor does not see the leaf entering a mouth.
The photographs and diagram in the paper make another constraint concrete. Sensor axes are fixed inside a device, so changing how the device sits on the head changes what its numbers mean. The male's data required an orientation adjustment because his harness placed the equipment differently. Consistent attachment is therefore part of the behavioural measurement, not a minor equipment detail.
A harder test of recognition
The team first used recordings from an individual to train a classification model, then asked it to label other recordings from that same giraffe. Their chosen method, a random forest, combined many branching decision rules to assign a behaviour from the motion measurements.
For these initial tests, performance was generally high. But the computer had already encountered the characteristic movements of the animal it was being asked to recognize. That resembles learning someone's handwriting and then reading another note from the same writer.
The second test withheld an entire animal. With the device used on all three giraffes, the model learned from two and was tested on the third. With the other device, the two females supplied the training and test records in turn. Only six behaviour categories shared sufficient observations for these comparisons.
This was a more relevant test of transfer to an unobserved giraffe. The published average category-accuracy scores fell from about 94% to 88% for one device and from about 90% to 82% for the other. These are the paper's averages of category-specific scores, including correct rejection of other behaviours, rather than a guarantee that those percentages of every activity in a wild day would be identified correctly.
The mistakes also had a pattern. Broadly similar movements—particularly rumination and feeding around intermediate head heights—were harder to separate. Feeding at ground level and walking were generally more distinctive in the comparisons between animals. The detailed results therefore matter more than a single overall score when choosing the question a sensor might answer.
Clean examples are only part of a real day
For model development, the researchers selected recording stretches containing just one behaviour and balanced the numbers drawn from each category. Mixed stretches were excluded. That made the relationship between movement and label easier to establish, but a wild giraffe can change activity midway through a recording.
A model trained on clean examples may force such a transition into an existing category. It may also mislabel a behaviour it never learned at all. Sleep-related observations, for example, came from the male on three recorded nights; the study did not demonstrate reliable sleep recognition across new animals. Posture-based observation also differs from measuring brain activity.
The equipment had practical limits alongside the classification limits. Real-time transmission from one device was interrupted by obstructions and sometimes by the giraffe lowering its head. Battery performance did not always match expectations. These are properties of the devices and settings tested at the time, not a verdict on every later tracking system.
A complementary study by Jason D. Wark and Katherine A. Cronin used trained human observers across 18 zoos to describe giraffe behaviour. Its standardized definitions and reliability checks emphasize the foundation on which automatic classification also depends: someone must first establish what the recorded behaviour means. A computer trained against uncertain labels would inherit that uncertainty.
The sensor study established a useful starting point for observing beyond a person's line of sight. Its next test is equally specific: more animals, realistic transitions and matched direct observations under field conditions. The goal is a behaviour record whose errors are understood well enough that a researcher can trust the biological conclusion drawn from it.
Sources
- Brandes, Stefanie; Sicks, Florian; and Berger, Anne (2021). “Behaviour Classification on Giraffes (Giraffa camelopardalis) Using Machine Learning Algorithms on Triaxial Acceleration Data of Two Commonly Used GPS Devices and Its Possible Application for Their Management and Conservation.” Sensors, 21(6), 2229. Full paper.
- Wark, Jason D.; and Cronin, Katherine A. (2024). “The behavior patterns of giraffes (Giraffa camelopardalis) housed across 18 US zoos.” PeerJ, 12, e18164. Related observation study.
