EweAcT: Ewe behaviour aligned to accelerometer data for activity monitoring in extensive grazing systems

arXiv:2608.09943 · cs.HC, cs.LG · Submitted 2026-07-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "EweAcT: Ewe behaviour aligned to accelerometer data for activity monitoring in extensive grazing systems".

Jane: The paper was written by Lucile Riaboff, Ny Aina Andriamampandry, Jean-François Bompa, Mathias Aletru, Christian Durand et al. from Université de Toulouse and INRAE and ENVT and UE0321 INRAE La Fage.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: Welcome back to the show, folks. Today we're looking at a paper that's got a title that's a mouthful: "EweAcT: Ewe behaviour aligned to accelerometer data for activity monitoring in extensive grazing systems." Jane, I have to say, the pun in the name alone made me smile.

Jane: It's a good one, right? EweAcT, like "ewe" and "act" combined. But the real story here is about putting little motion sensors on sheep and watching what they do all day. We're talking about one hundred twenty Romane ewes, wearing necklaces with accelerometers, basically like a Fitbit for sheep.

Tom: And I love that they're not just doing this in a barn. These are sheep living on two hundred eighty hectares of rangeland in southern France, out in the heat, on slopes, in the middle of an extensive grazing system. That's a pretty tough environment to be collecting data in.

Jane: Exactly. And the goal is to build a library of data so that eventually, we can just look at the accelerometer signal and know what the sheep is doing without having to watch hours of video. Grazing, ruminating, resting, moving—those are the big ones they're tracking.

Tom: So we're essentially teaching a computer to read sheep body language from a little sensor on their neck. That's the dream, anyway. And this dataset is the raw material to make that happen. I'm excited to dig into how they actually pulled this off.

Jane: Me too. And the fact that they've made it publicly available is huge. It means any research group in the world can start working on this problem without having to spend months out in a field with a camera. That's a real gift to the community.

Tom: Absolutely. It's one thing to have the idea, but to actually release the data so others can build on it? That's how science moves forward. Let's get into the details of what's actually in this dataset.

Summary: Tom: So Jane, we've got the title, we've got the general idea. But what's the actual summary of this paper? What did they really accomplish here?

Jane: Well, Tom, they've created a dataset called EweAcT, and it's essentially a massive, carefully curated collection of accelerometer readings that are perfectly synced up with video footage of what the sheep were actually doing at that exact moment. We're talking about seventy-nine hours of data, with over seven million individual readings.

Tom: Seven million readings. That's not a small sample. And each one of those readings is tagged with a behaviour—grazing, ruminating, resting, moving, or a catch-all "other" category.

Jane: Right. And the clever part is how they aligned the sensor data with the video. They built this little spinning cylinder that holds ten accelerometers at a time. When they spin it, it creates a very distinct pattern in the sensor data. They know exactly when they spun it, and that lets them sync the sensor clock with the video clock.

Tom: It's like a clapperboard for accelerometers. You know, the thing they use in movies to sync the camera and the sound. They created a physical signal that both systems can see, and that's how they line everything up.

Jane: Precisely. And they did this at the start and the end of each recording session to correct for any drift in the sensor's clock over time. That's the kind of meticulous detail that makes this dataset reliable. It's not just a pile of numbers; it's a carefully engineered resource.

Tom: And they didn't just do this in one field, on one day. They collected data in March, June, July, and even November, covering different seasons and different conditions. Some paddocks were flat, some were sloped. Some days were cool, some were scorching hot.

Jane: That diversity is key. If you want to build a model that works in the real world, you need to show it the messiness of the real world. A sheep moving on a flat pasture looks different from a sheep moving on a slope, and a sheep that's panting in the heat looks different from one that's comfortable.

Tom: So they've really thought about making this dataset representative of the challenges of an extensive system. It's not just a lab experiment. It's a real-world test. And the fact that they've got data from sheep of different genetic lines adds another layer of complexity and usefulness.

Jane: It does. And that's what we should talk about next—how they're using this to understand the animals themselves, not just the technology.

Improvements: Tom: So Jane, we've talked about what the data is. But what does this paper suggest we can actually *do* with it? What's the improvement over what we have now?

Jane: The big improvement is moving from watching sheep to understanding them at scale. Right now, if you want to know how a sheep is coping with heat stress, you have to watch it. You have to sit there and observe. With a dataset like this, you can train a model to do that observation for you, continuously, twenty-four/seven.

Tom: So instead of a human watching a video, we have an algorithm that can look at the accelerometer data and say, "Hey, this sheep is panting," or "This sheep is moving a lot more than usual." That's a huge leap forward.

Jane: And the paper specifically mentions that they collected data during a heatwave in July. They saw the sheep panting, they saw them changing their behaviour. That's the kind of extreme event we need to understand better, especially as our climate gets more unpredictable.

Tom: Right, and this connects to the genetic lines they mentioned. They have sheep selected for high sociability and low sociability, high docility and low docility. The idea is to see if these different genetic backgrounds respond differently to stress. Do the more sociable sheep cope better with heat? Do the more docile ones handle human interaction better?

Jane: That's the exciting part. This dataset isn't just about building a better activity monitor. It's about using that monitor to answer big biological questions. Questions about animal welfare, about adaptation, about how genetics influence behaviour in real-world conditions.

Tom: And that has practical implications for farmers too. If you can automatically detect that a sheep is showing signs of distress, you can intervene earlier. You can provide shade, you can provide water, you can change your management practices.

Jane: Exactly. It's about giving farmers better tools to care for their animals. And the authors are clear that this is a first step. They've created the foundation, and now they're inviting the rest of the world to build on it. That's the real improvement—the opening of a door.

Tom: A door to a future where we can keep a constant, watchful eye on the health and well-being of every single animal in a flock, even when they're spread out over hundreds of hectares. That's pretty powerful.

Jane: It is. And it all starts with getting the data right, which is what they've done here. Let's look a bit closer at the first page of the paper to see how they set the stage for all of this.

First Page: Tom: So Jane, we've been talking about the big picture. Let's zoom in on the very beginning of the paper. What's the hook? What's the problem they're trying to solve?

Jane: The first page sets up the context beautifully. It talks about agroecology and the push towards extensive systems where animals are raised outdoors. That's great in theory, but it means the animals are exposed to a lot more variability—weather, predators, resource scarcity.

Tom: And that variability can hit their health and welfare. The paper mentions things like heat waves and parasitism as examples of environmental perturbations. So the question is, how do we keep an eye on these animals when they're out there in the wild?

Jane: And that's where behaviour comes in. Behaviour is a window into how an animal is coping. If a sheep is restless, if it's not grazing normally, that's a sign something might be wrong. But you can't have a person watching every sheep all the time.

Tom: So they need an automated system. And to build that automated system, they need data. That's the core need this paper addresses. They're filling a gap by providing high-quality, annotated accelerometer data specifically for sheep in these extensive conditions.

Jane: Right. And they point out that this is, to their knowledge, the first dataset specifically designed for this purpose. There's data on cows, there's data on sheep in more controlled settings, but not this combination of extensive conditions, genetic diversity, and environmental challenges.

Tom: And they're not just throwing data out there. They're also providing the tools they used to create it—the Python notebooks, the metadata, the sample files. They want people to understand exactly how the data was made, so they can trust it and even extend it.

Jane: That transparency is so important. They've thought about the end user. They've thought about the person who's going to try to train a model on this data and might run into questions. They've tried to answer those questions in advance.

Tom: So the first page really lays out the "why." Why this dataset matters, why it's needed, and why it's trustworthy. It's a strong foundation for everything that follows.

Jane: It is. And it makes me think about the future. Where does this go from here? Let's wrap up our thoughts on that.

Conclusion: Tom: Well, we've spent a good amount of time with "EweAcT: Ewe behaviour aligned to accelerometer data for activity monitoring in extensive grazing systems," and I think it's safe to say this is a really solid contribution.

Jane: It really is, Tom. We've got a publicly available dataset with seventy-nine hours of carefully annotated accelerometer data from one hundred twenty ewes, covering the main behaviours of grazing, ruminating, resting, moving, and other. And they've done the hard work of syncing the sensors with the video, so the labels are reliable.

Tom: And the conditions are diverse—slopes, heatwaves, different seasons. That's what makes it useful for real-world applications. It's not a sterile lab dataset.

Jane: The potential impact is huge. This could lead to automated systems that help farmers monitor the well-being of their flocks continuously. It could help researchers understand how different genetic lines respond to stress. It could even help us breed more resilient sheep for the future.

Tom: And it's a great example of open science. They've shared the data, the code, and the methodology. They're not keeping it to themselves. They're inviting the whole research community to build on their work.

Jane: There are limitations, of course. The dataset is mostly young ewes of one breed, and there's a lot of data from hot conditions. But they're upfront about that, and they see it as a starting point for future data collection.

Tom: So, as we say goodbye to this paper, I think we can say it's a job well done. It's a building block for a smarter, more caring approach to animal agriculture.

Jane: Absolutely. And we're excited to see what models people build using this data. For now, that's our take on EweAcT. Thanks for listening, and we'll catch you on the next one.

Tom: Take care, everyone.

Lucile Riaboff, Ny Aina Andriamampandry, Jean-François Bompa, Mathias Aletru, Christian Durand, Sébastien Douls, Gaëtan Bonnafe, Morgane Costes-Thiré, Guillaume Delosières, Jean-Marc Mongrelet, Enzo Niro, Némuel Tadi, Séverine Deretz, Sara Parisot, Margot Lamarque, Dominique Hazard, Emilie Cobo

Université de Toulouse · INRAE · ENVT · UE0321 INRAE La Fage

cs.HC, cs.LG

Submitted: 2026-07-01

Updated: 2026-08-12

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 58/100

The gist: low and high social attractiveness, noted S- and S+, and low and high tolerance towards humans, noted H- and H+." They were reared under the extensive system applied to the Experimental Unit of La

Terminology

Summary

Summary

The paper introduces the EweAcT dataset, which is specifically designed to monitor sheep behaviour in extensive systems from accelerometer data, encompassing a wide range of conditions, such as sloping paddocks, different seasons and varying animal thermal states. The dataset includes 79 hours of tri-axial accelerometer data aligned with behaviours manually annotated from video recordings for 120 Romane ewes born between 2021 and 2024. The ewes were derived from two divergent genetic lines after three and four generations of selection started 10 years ago: low and high social attractiveness, noted S- and S+, and low and high tolerance towards humans, noted H- and H+. They were reared under the extensive system applied to the Experimental Unit of La Fage (UEF, INRAE, Saint-Jean-et-saint Paul, Aveyron) where 250 sheep were reared exclusively outdoors on 280 hectares of rangeland in southern France.

The dataset was collected in two batches. The first batch was collected on March, June and July 2024 at the UEF under a range of extensive conditions, including sloping pastures and heat-wave periods. Ewes were "equipped with accelerometer neck-collars specifically designed for young sheep on pasture. They were grouped on experimental paddocks for 4 to 8 hours and provided with fresh grass and ad libitum access to water. The animals were simultaneously video-recorded using an elevated CCTV camera. Behaviour annotation was carried out using Behavioral Observation Research Interactive Software (BORIS), focusing on the main behaviours on pasture: Grazing, Ruminating, Resting, Moving, and Other, which groups all remaining activities. Annotations and corresponding accelerometer sequences were aligned using Python language, based on a time synchronization procedure. The second batch was acquired on November 2025 to supplement the dataset with the moving activity. For that purpose, ewes were equipped with the accelerometer collars and moved on tracks from the housing area to the pastures, corresponding to an approximately 10 minute-walk. The start and end times of the moves for each ewe were used to align the corresponding accelerometer data with the moving activity. These data were then merged with the dataset from the first batch."

The dataset is a table saved as a PARQUET file (EweAcT.parquet; 62.5 Mo) containing 9 columns and 7,108,285 rows. Each row represents an accelerometer reading along three dimensions (AccX, AccY, AccZ) collected from one animal at a specific timestamp (sampling rate: 25 Hz) and aligned with a behaviour. The behaviours are Grazing, Resting, Ruminating, Moving, and Other, with definitions provided in a table. The dataset also includes a metadata table (metadata EweAcT.xlsx) with key characteristics for each ewe, including genetics, year of birth, month of observation, location, slope, thermal status, and average outdoor temperature. A Jupyter Notebook (AX3BehaviourAlignment.ipynb) is also provided to illustrate how the dataset was constructed.

Time synchronization was applied to the first batch of data to align accelerometer timestamps with video timestamps to a common UTC clock. This was achieved by generating a distinct pattern in the accelerometer signal using rotating cylinders, with the exact time saved on the Raspberry Pi's UTC clock. The accelerometer time series was corrected for time offset and drift. For the second batch, accelerometer data corresponding to each move were identified using start and end times recorded with a smartphone clock, and sequences were carefully inspected to remove instances of other activities.

The paper notes limitations: the animals were relatively homogeneous (only female Romane sheep, mostly ewe lambs), and much of the Resting and Ruminating data were collected under hot conditions (above 20°C) with heat-related behaviours. The authors state that the dataset is ready to use for applying artificial intelligence models to classify the 5 main behaviours of sheep under extensive grazing systems from accelerometer data.

Improvements for AI systems

Based on the scientific paper, here are the specific improvements I can make to AI systems and what the improved system can do:

  • Improvement: Train a deep learning model (e.g., CNN, LSTM, or Transformer) specifically for classifying sheep behaviours from raw tri-axial accelerometer data at 25 Hz, using the 79 hours of labelled data from 120 ewes.

  • What the improved system can do: Automatically classify five behaviours (Grazing, Ruminating, Resting, Moving, Other) in real-time or near-real-time with high accuracy, replacing manual video annotation which took observers several months.

  • Improvement: Incorporate the metadata (slope, temperature, thermal stress, genetic line, season) as auxiliary inputs or use domain adaptation techniques to make the model invariant to these conditions.

  • What the improved system can do: Maintain high classification accuracy across sloping pastures, heat-wave periods (>20°C), different seasons (March, June, July, November), and for animals from different genetic lines—critical for deployment in real extensive grazing systems where conditions vary widely.

  • Improvement: Apply class-weighted loss functions, oversampling, or synthetic data generation (e.g., SMOTE) to address the imbalance shown in Figure 1 (e.g., Moving is underrepresented in batch 1, while Grazing dominates).

  • What the improved system can do: Avoid bias toward dominant behaviours, ensuring rare but critical behaviours like Moving and Other are detected reliably, which is essential for detecting distress or abnormal events.

  • Improvement: Use sequence-to-sequence models with sliding windows (e.g., 5–30 seconds) to capture temporal dependencies, since behaviours like Ruminating have chewing interruptions <10 sec and Grazing has head-down periods.

  • What the improved system can do: Distinguish between similar accelerometer patterns (e.g., standing still vs. resting while standing) by leveraging temporal context, reducing false positives.

  • Improvement: Pre-train a base model on EweAcT, then fine-tune on smaller datasets from other breeds or systems (e.g., adult sheep, different terrains).

  • What the improved system can do: Rapidly adapt to new environments with minimal labelled data, reducing the need for months of annotation in each new setting—saving significant time and cost.

  • Improvement: Train an autoencoder or one-class classifier on normal behaviour patterns (Grazing, Ruminating, Resting) to detect deviations (e.g., panting, agitation, abnormal transitions).

  • What the improved system can do: Alert farmers in real-time to signs of heat stress (panting), distress, or abnormal behaviour, enabling early intervention for animal welfare—a key goal of the paper.

  • Improvement: Use the annotated sequences to train a higher-level model (e.g., HMM or Transformer) that predicts behaviour transitions and daily activity budgets.

  • What the improved system can do: Provide insights into how ewes allocate time to different activities across the day, and detect shifts in behaviour patterns due to environmental perturbations (e.g., heat waves, predator presence), supporting studies on animal adaptation.

  • Improvement: Use the time-synchronization procedure (pattern detection via signal.find peaks) to develop a calibration module that corrects for timestamp drift and sensor misalignment in new deployments.

  • What the improved system can do: Automatically align accelerometer data from multiple sensors to a common clock, ensuring that predictions are correctly timestamped even after hours of deployment—critical for longitudinal monitoring.

The improved AI system can:

  • Continuously monitor individual sheep behaviour in extensive grazing systems without human intervention.

  • Detect heat stress and other welfare issues in real-time, enabling timely intervention.

  • Generate behaviour budgets over days/weeks to study genetic differences in sociability and docility.

  • Adapt to new conditions (breeds, terrains, seasons) with minimal additional data.

  • Provide reliable, timestamped data for downstream analysis, even in remote field settings.

These improvements directly address the paper's stated goal: enabling AI models to monitor sheep behaviour in extensive systems, supporting studies on animal adaptation to agroecological challenges.

Related papers