The Bat No Bat Sorter

The Problem: A Data Deluge

Our project faces one major challenge: scale. A single AudioMoth recorder can generate gigabytes of data every night. On a rainy or windy evening, our full network of 100+ recorders could produce over 3.2 terabytes of raw audio. That’s thousands of hours of sound, and the vast majority is just background noise like wind, traffic, or insects. It is simply impossible for our team to listen to it all. We needed a smart, automated filter to find the "needle in the haystack."

Our Solution: An AI Built by Hand

The "Bat No Bat Sorter" wasn't built overnight. It’s the result of an iterative "bootstrap" strategy, where each step made the next one possible.

Step 1: We first built a simple program based on the natural characteristics of echolocation calls to find potential bat calls.

Step 2: Our human experts manually sorted thousands of these sound snippets to create a "gold standard" dataset.

Step 3: We used this set to train our first AI, the "Alpha Model".

Step 4: We then used that model to sort more data, and our team refined its results again, manually verifying over 52,400 new snippets.

Step 5: This large, human-verified dataset was used to train the final, powerful "Beta Model". This is the engine inside the sorter you use today.

This means our AI is powerful and trustworthy because it was trained by our own experts on real-world data collected right here in Moorhead. In fact, we are yet to find it has made a mistake in identifying a data file as bat or no bat!

How It Works: Your On-Site Data Lab

When you run the "Bat/No-Bat Sorter" (part of our "BatFieldLauncher" suite) , you are running a sophisticated scientific analysis tool right on your own laptop, no internet or cloud resources required. Here’s what happens:

• Scan: The tool instantly scans your audio file in 0.5-second chunks.

• Visualize: It turns each tiny chunk into a spectrogram—a "fingerprint" image of the sound from 18 to 80 kHz.

• Predict: Our trained "Beta Model" looks at that image and predicts the probability that it contains a bat call.

• Verify: It's not just the AI. The program adds a final "common sense" check, making sure there's enough bat activity (e.g., at least 1.5 seconds) to be confident.

• Sort: The sorter then tags your file as "Bat," "Noisy Bat," or "No Bat," placing it in the correct folder.

Why It Matters: Turning Noise into Science

This tool is what makes a project of this scale possible. The "Bat No Bat Sorter" has demonstrated exceptional performance and near-perfect precision in real-world field testing. It allows us to confidently filter terabytes of raw data down to a small, information-rich fraction only containing confirmed bat activity.

This system transforms "data overload into scientific opportunity". When you run this sorter, you are not just a data collector. You are performing the first and most critical step of our scientific analysis, enabling discoveries that would otherwise be buried in noise.