BirdNetPi: false positives & %conf level

 I don't like false positives 

...but I don't want to miss anything!

That's the dilemma when setting the BirdNetPi %confidence level.

Setting %conf too low is bound to give you false detections (i.e. incorrectly identified bird species). But then setting the %conf too high may mean you miss perfectly correct species!

I thought I'd take another look at this, as my upgraded BirdNetPi has now been running for about 5 months.

With improvements to the microphone wiring, mounting, positioning and also an upgrade from the McGuire to Natchzuster BirdNet-Pi software, the system is reporting higher numbers of both detections & species for 2025 than it was for 2024. But just like last year, some recordings have been incorrectly classified.

These include the American black-crowned night-heron (I think this was due to New Years fireworks!), nightingale (which I think was a blackbird), raven (which sound just like our neighbours dog), red-legged partridge & ring-necked pheasant (...again, the dog's in the frame!) and turtle dove (which are probably a mix of collard doves & wood pigeon).

So these species will be disregarded.

 

%confidence level

In addition, I exported the database to a CSV file and then filtered the results to assess the effect of changing the %confidence level from 75% to 80%;

 - with a total of 189,910 results, 48,650 fell within the range 75% - <80%

 - this represents approx 20% of total detections

 - this includes 9 species that would not be counted for an 80% confidence level

 - these excluded species are:-

•   greenshank (1 detection)
•   common tern (1)
•   peregrine falcon (1)
•   rock pigeon (1)
•   tree sparrow (1)
•   mistle thrush (2)
•   red-legged partridge (2)
•   ringed plover (2)
•   mute swan (3)

My conclusion is that these species detections in the 75 - 80% range are suspect, and therefore I'm happy to exclude them.

But for consistency, I won't adjust %conf at the moment, I will only change it at the start of 2026.

There a few species in the database from the last 2 weeks of December 2024. So once they are also removed, my unique species count for 2025 so far is 94.

Other BirdNet related posts: https://captainbodgit.blogspot.com/search/label/BirdNET