First Thoughts on Confidence Levels and A Grackle Migratory Group

By Chuck Hennet — September 10, 2025

One thing I wrestle with on this project is how to handle the confidence scores that BirdNET gives each detection. Every time the system hears something, it tells me how sure it is — from 0 (not at all sure) to 1 (completely certain).

Some people might think it’s a mistake to include every detection in the database, even the ones with very low confidence. It could look like I’m adding a lot of noise, or just trying to boost the detection numbers. But I think including everything — even the long shots — gives a fuller picture of what might be out there.

This morning, while walking the dogs, I saw a group of migrating grackles — maybe thirty birds — moving through about 50 yards from Station 3’s microphone. When I checked the data later, BirdNET had detected them: 24 separate entries over a ten-minute window between 7:26 and 7:36 a.m. The confidence levels were low — the highest at 0.3018, most between 0.10 and 0.30 — but the match was there.

Including low-confidence detections matters. Sometimes, like with the grackles this morning, a low-confidence detection lines up with real birds seen on the farm — a visual confirmation that backs up the data. Other times, those low numbers might just hint at a possibility, a layer of probability in the background. But that’s part of what makes this project interesting: it’s not about certainties, it’s about patterns, potential, and paying attention over time.