Terracorder: Sense Long and Prosper

Josh Millar, Sarab Sethi, Hamed Haddadi, Anil MadhavapeddyarXiv 2024 · DOI · arXiv

Abstract

Biodiversity monitoring in remote, off-grid environments requires a delicate balance between sensor activity and power conservation. Traditional methods rely on fixed schedules or power-hungry "always-on" systems, both of which are suboptimal for capturing rare ecological events. This paper introduces Terracorder, an ultra-low-power multi-sensor platform that employs on-device reinforcement learning (Q-learning) to autonomously adapt its sensing schedule to local environmental patterns. By learning when activity is most likely, Terracorder achieves an 85.3% event capture rate while extending battery life to over one year—a 55.1% improvement over standard fixed-interval baselines.

Introduction

Ecologists face a significant challenge when monitoring wildlife in remote areas like rainforests: the lack of power infrastructure. Current solutions often fall into three categories: fixed schedules that miss events or waste energy, always-on devices that drain batteries rapidly, or triggered sensors that are limited in the types of wildlife they can detect. Terracorder addresses this by "learning" the rhythm of its environment, waking up more frequently during high-activity periods (e.g., the "dawn chorus") and remaining in deep sleep during "dead zones."

Methods

The Terracorder hardware is built around the PowerFeather module (ESP32-S3), optimized for a deep-sleep current of approximately 19µA. The system integrates a 5MP camera, an I2S microphone, and a PIR motion sensor.

Terracorder hardware components including the PowerFeather module, 13400mAh battery, and various sensors.

To manage power, we implemented a model-free Q-learning algorithm on-device. The scheduler treats the time of day as a state and the activation frequency as an action. It receives "rewards" for successful detections—determined via low-power Goertzel filters or TensorFlow Lite Micro CNNs—and "penalties" for activations that yield no data. This allows the device to fine-tune its behavior without human intervention.

Results

Evaluation using acoustic data from the SAFE Project in the Borneo rainforest shows that the Q-learning (QL) scheduler significantly outperforms fixed-interval schedules.

Comparison of Captured Events (%) vs. Battery Lifetime (years) for various fixed schedules and the Q-learning (QL) approach.

As shown in the performance analysis, the QL approach (represented by the star) provides a superior balance between data capture and longevity. It captured 85.3% of events with an estimated battery life of 1.07 years using a 13400mAh battery. In contrast, a 3-second fixed schedule captures nearly all events but lasts only 0.69 years, while longer fixed intervals significantly sacrifice data quality for marginal gains in longevity.

Discussion and Limitations

While Terracorder shows great promise, challenges remain. Simple frequency filters can be triggered by environmental noise like wind, potentially misleading the RL agent. Furthermore, the agent requires an initial learning period, or "cold start," before it becomes fully optimized. Future work will focus on collaborative sensing, where networked devices "ping" each other to reduce redundant recordings, and improving physical durability for multi-year deployments in harsh tropical climates.

Conclusion

Terracorder demonstrates that reinforcement learning is not only feasible on microcontrollers but is a powerful tool for ecological conservation. By aligning sensor activity with the natural rhythms of the environment, we can achieve long-term, high-fidelity monitoring on a minimal power budget.

Summary updated

BibTeX
@misc{millar2024terracorder,
  author = {Josh Millar and Sarab Sethi and Hamed Haddadi and Anil Madhavapeddy},
  title = {{Terracorder: Sense Long and Prosper}},
  year = {2024},
  eprint = {2408.02407},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.02407},
}
Figure