This case study presents an alternative route for studying avian incubation behaviour. While camera traps and visual recognition models dominate much of the conservation technology landscape, they are not always viable, especially in remote or densely vegetated ecosystems. In such settings, sensor-based approaches can offer a powerful, scalable, and less intrusive solution.
Here, researchers in Colombia (@jcguerra10, @bmgarrido, and me) introduce a modular, open-source system that combines environmental sensors with a deep learning model to detect incubation behaviour in wild birds. Instead of manually labelling thousands of hours of footage, our method uses simple temperature loggers to identify critical behaviours—when a parent bird enters or leaves the nest (on- and off-bouts) and when it initiates nocturnal incubation.
The core of the system is a Bidirectional Long Short-Term Memory (BiLSTM) model, trained on over 2,200 nest-days from more than 120 Neotropical bird species. The model was developed following a three-step validation process: an initial 80/20 training-test split, followed by five-fold stratified cross-validation, and a final evaluation on an independent dataset. This robust methodology ensured reliable generalisation and performance, with theoretical error rates below 5%, based on macro F1 scores consistently around 0.95. These results suggest the model can accurately classify key incubation behaviours—on-bouts, off-bouts, and nocturnal incubation—even across species and ecosystems not included in the training data.
Exploratory Findings
In colaboration with two projects Manu and Gradientes Colombia (leadered by PhD Gustavo Londoño), researchers also tested this method to study incubation strategies in Myioborus miniatus and M. melanocephalus, species distributed along Andean gradients. The model processed 328 full nest-days from both warbler species in just 28 seconds, with a classification accuracy exceeding 99%.This approach is not only accurate—it’s transformative. A task that used to take days or weeks of human effort is now fully automated, making large-scale behavioural analysis feasible, even in data-limited contexts.
For decades, analysing bird incubation behaviour from nest temperature or camera traps data has been a slow and labour-intensive process. Early methods with nest temperature required biologists to manually review spreadsheets, spending up to eight hours just to annotate a week's worth of data for a single nest. Even with the advent of tools like RAVEN, incR, or NestIQ, analysis times remained lengthy—up to eight hours per 15 days of data—and still depended on manual adjustments and calibration. This not only made large-scale studies impractical but also introduced room for human error. These challenges highlighted a clear need for faster, more reliable tools capable of handling diverse datasets without the burden of constant fine-tuning.
How to Use It
The trained model is available on GitHub:
👉 GitHub Repo
- You can load your data and perform inference in PyTorch using the
.pthfile at the final model. - Basic instructions and examples are included in the repository.
Why It Matters
By eliminating the need for complex visual data, this sensor-driven approach makes behaviour modelling more accessible, especially in the tropics. It lowers the cost, increases the scalability, and preserves the privacy of sensitive ecosystems. The use of open-source tools also ensures that other researchers can adapt and improve upon this work, including for new behavioural traits or additional sensors.
_______________________________________________________________________________________________________
🦉 Coming soon: the model will be available as a PyPI package with a Dash-based web interface, enabling use with zero coding required.
Let us know how you’d like to apply it in your own work—or if you want to contribute to future versions!
_________________________________________________________________________________________________________
Caveats
The model's performance is fundamentally tied to its training data. Accuracy may decrease when applied to species, climates, or sensor configurations significantly different from our Neotropical dataset. The model's reliance solely on temperature can be a limitation in environments where ambient temperatures are high, obscuring the thermal signatures of off-bouts. Future work should aim to incorporate complementary data streams from pressure to motion sensors, for which our modular architecture is well-suited. To enhance its impact and accessibility, our immediate goal is to package BoutScout into a user-friendly tool.
We thank Universidad Icesi and PhD Gustavo A Londoño who guided us throughout this journey—for their unwavering support, mentorship, and belief in the potential of this work. Also a thanks to Angie Pamela López and Mariana Quintero, both experienced conservation biologists, provided invaluable assistance in labelling incubation behaviours across over 200 nests, ensuring the quality of the training dataset.
Add the first post in this thread.