v2.4 List model anomalies

[Mattk70]

Hi @kahst @Josef-Haupt ,

For a while now, there have been list anomalies that have been troubling me. This is using the list model in 2.4, discussed here: https://github.com/kahst/BirdNET-Analyzer/discussions/234

Here's an example - two species that seem wrong for the list at my location: Corn Bunting (scarce generally but not unknown and locally not uncommon) and Arctic Skua (Parasitic Jaeger) - an extremely rare vagrant locally.

Location: Lat 51.9, Lon -0.4 Use week, no

Parasitic Jeager: present in the list when the species threshold is 0.07 excluded at 0.08 Corn Bunting: not present at 0.01 (in fact the bird is not included until the threshold drops to 0.004)

I thought this may be to do with some odd eBird checklist reporting, so I checked eBird for my region.

Corn Bunting is recorded regularly, if not frequently: https://ebird.org/barchart?r=GB-ENG-BDF&bmo=1&emo=12&byr=2010&eyr=2024&spp=corbun1

The Arctic Skua has not been recorded at all: https://ebird.org/barchart?r=GB-ENG-BDF&bmo=1&emo=12&byr=2010&eyr=2024&spp=parjae

There are other species I could draw attention to, but hopefully this is a good enough illustration. If there is a theme, a systematic error, it is that seabirds (auks, skuas and petrels in particular) are considerably more prevalent in the list than one would expect.

Do you have any ideas why this is?

Thanks!

[alexbelgium]

Hi, thanks I have the same observation especially for night birds such as owls.

For example are completely improbable in my region (as a forest far from sea): Morus bassanus_Fou de Bassan - 0.009305058 Pandion haliaetus_Balbuzard pêcheur - 0.009156795

And the very commun Strix Aluco (nearly the only hit every night) has a lower value than a Morus bassanus! Strix aluco_Chouette hulotte - 0.008530801

The Athene noctua is also a very rare bird but at the same frequency as the Strix alluco Athene noctua_Chevêche d'Athéna - 0.00852739

[AxelHo]

Hi same for me. i am just testing the server version. My testfile is 100% Strix Aluco

Result without lat and lon [ 'Strix aluco_Waldkauz', 0.9916 ],

Result with "lat": 52, "lon": 8 here is a Strix aluco near our house it is not in the list of 10 results.

[alexbelgium]

I had to set threshold of detection to 0.0007 to include all expected species ; which allows lots of "abnormal" species but still removes 3 quarters of the full list so I consider worth it. This allows to show all prey birds (night and day).

[AxelHo]

Thanks, I will give it a try. For now i switched Back to V2.3 without setting the lat and Long. The results are good, some exotics but Not greater than my 70% Filter. Next I want to Test the chirpity model, on the paper ITS faster than Bird-Net Analyzer. And I know JS better than Python 😁

[Mattk70]

@AxelHo, Chirpity uses the same list model as BirdNET, so you'll get the same result. There are other ways to adapt the species detected in Chirpity, The only sure fire way right now to get the species you want is to use a custom list.

[Mattk70]

@kahst @Josef-Haupt - I noticed you’ve been triaging old issues. Did you look into this one? It'd be great to hear your thoughts on the reason for the seabird bias even if a fix is intractable.