Autonomous Monitoring of Insects (AMI)

Background

We are currently facing a major environmental crisis: a 75% drop in flying insect populations over recent decades. Because insects are vital for pollinating crops and keeping soil healthy, losing them threatens entire ecosystems. To protect them, scientists need a continuous, widespread way to track insect populations in the wild.

Traditionally, tracking insects meant catching them by hand and identifying them one by one—a process that is slow, exhausting, and difficult to do on a large scale. While modern automated camera traps help by taking high-quality photos, they have a big flaw: they simply save thousands of massive files to memory cards. Analyzing those photos later requires expensive, power-heavy computers and manual internet uploads, meaning there is a massive delay before scientists know what is happening in the forest.

Project AMI solves this by building a mini-AI "brain" directly into the insect trap. Instead of just saving photos or sending them to the cloud, the trap identifies the insects on the spot using a tiny fraction of the power a standard lightbulb uses (less than 3 watts), focusing specifically on important East African moth families.

Accomplishments

We have successfully shifted the project from a passive camera setup into an intelligent, self-contained sorting system. Our core achievements include:

• A Massive Photo Library: We have gathered a foundational collection of 100,000 snapshots of nocturnal insects. This local field imagery is combined with global wildlife databases to teach our smart software exactly what different species look like.
• Automatic Photo Cleaning: We built a smart pipeline that automatically cleans and prepares the photos. It automatically crops the images to focus perfectly on the insect and removes background noise before using them to train the AI.
• A Custom AI Program ("MothNet"): We designed a highly specialized, lightweight AI program called MothNet. By changing how the software processes visual data, we made it 8.5 times faster and lighter than standard AI programs, allowing it to run on simple hardware.
• Squeezing AI onto Tiny Microchips: Standard AI requires massive computers, but we used advanced shrinking techniques to teach a microchip-sized version of our program to mimic the large one. We trimmed away useless data and simplified the math, making the final AI program 4 times smaller (under 1 megabyte) so it easily fits on a tiny chip.

Next Steps

Our next phase focuses on moving from the laboratory into active, long-term forest protection:

1. Microchip Testing: We will run intensive tests on the physical microchips using power meters to ensure the software remains incredibly fast and energy-efficient under simulated wilderness conditions.
2. Real-World Forest Trials: We are taking these smart traps into the wild at the DeKUT Conservancy to prove that the AI can successfully identify live moths in real-time.
3. Making it Affordable and Open Source: We are finalizing the software to be entirely open-source and free for public use. This keeps the entire hardware setup highly affordable, making advanced conservation tools accessible to researchers globally.
4. Adding Sound Sensors and Long-Range Radios: In the future, we plan to add low-power microphones to identify insects by their sounds alongside their photos. We will also integrate long-range, low-power radio systems (LoRaWAN) to beam simple population updates out of deep, dense forests where there is absolutely no cell phone signal.

Publications