This TinyML Project Listens Out for Dangerous Mosquitoes Using Low-Cost Hardware and Edge Impulse

Researchers from the University of Khartoum, the International Center for Theoretical Physics, and the Universidade Federal de Itajubá have detailed a novel approach to monitoring disease-carrying mosquitoes — by using a low-power tinyML system to track dangerous mosquitoes based on the sound of their beating wings.

"Acoustic detection of mosquitoes has been studied for long and ML [Machine Learning] can be used to automatically identify mosquito species by their wingbeat," the team explains of its work. "We present a prototype solution based on an openly available dataset, on the Edge Impulse platform and on three commercially-available tinyML devices. The proposed solution is low-power, low-cost and can run without human intervention in resource-constrained areas. This insect monitoring system can reach a global scale."

The idea is simple: tracking the presence of mosquitoes by listening out for the sound of their beating wings, but in a way that doesn't require expensive equipment, human intervention, or a reliable power supply and that can distinguish harmless flying insects from those likely to carry diseases. To deliver this, the team turned to three off-the-shelf devices with tinyML-capable microcontrollers and embedded microphones: the Arduino Tiny Machine Learning Kit (Nano 33 BLE Sense), the Arduino Portenta H7, and the Seeed Studio Wio Terminal, all configured to communicate their findings via a low-power long-range LoRaWAN connection.

The tinyML model itself was based on a public dataset collecting 20 different mosquito wing-beat sounds, from which the team selected two of the most dangerous types: Aedes aegypti and Aedes albopictus. The audio recordings — including recordings of other mosquito species and background noise with no mosquitoes at all — were then pre-processed and fed into Edge Impulse Studio for training and testing.

The resulting model, tested at 94 percent accuracy with the test dataset, was then optimized using quantization in TensorFlow Lite and the Edge Impulse EON Compiler to reduce the memory required for operation — making it suitable for deployment to the low-cost devices chosen, which in the field would run from an external battery "for several days" with the option of a solar panel for longer runtimes.

"Several other solutions based on ML and audio sensors have been presented," the team admits, "but the uniqueness of the proposed system is that is adds field-readiness to accurate ML-based categorization. We have shown that commercially available tinyML devices can be used for a week in the field to monitor the proliferation of mosquitoes. As future work we plan to optimize power consumption by developing our bare-bone embedded device to only measure audio signals and send results via LoRaWAN."

The team's work has been published in the Proceedings of the 2022 ACM Conference on Information Technology for Social Good; an open-access preprint copy is available on GitHub, along with the project source code under the reciprocal GNU General Public License 3.