Some Problems with mi station Piscis #4960

[jimenoma]

Hi there,
I recently installed BirdNetPiPy; previously, I was using the Nachtzuster version of BirdNetPi.
My station is #4960, "Piscis," located in Navalafuente, Madrid, Spain.
The behavior is different now: it identifies a much higher absolute number of birds, but far fewer species.
Previously, my station ranked among the top 20 or 30 on the BirdWeather.com "Top Stations" list; now it is in the top 5—for instance, today it is ranked second.
However, I used to identify between 20 and 30 different species daily, whereas now the figure hovers around 10–13 (I’ve included screenshots).
As you can see from the screenshot, I believe the detection model being used is the same one I used in my previous BirdNetPi installation, and the settings are the defaults.
I need to know if I should modify the settings to once again identify species that were previously detected (currently, Common Buzzards—Buteo buteo—and Red Kites—Milvus milvus—do not appear).
Another issue concerns the BirdWeather app map: my station apparently isn't being detected there, even though the bird count seems to be updating in the database. (In the map screenshot, the bottom-right section shows 0 stations, 0 detections, and 0 species, yet the top-center area displays "12 species | 15,205 detections"; whereas with the Nachtzuster BirdNetPi version, the station was visible and showed "1 station," etc.)
Sorry for all these questions, but before proceeding with your excellent project, I wanted to clarify these points.
Many thanks and best regards from Spain,
MarcoA.

[Suncuss]

Hi MarcoA, thank you so much for the report.

I checked the BirdWeather site for #4960.

Seems like there is a mis match between the station coordinate and the coordinate of the latest detection.

Could you check if the coordinate you set in the BirdNET-PiPy is correct? seems to be missing a zero for the latitude.

[jimenoma]

Hi Suncuss:
I have changed to the correct values, before mi station was in the North Pole. JEJE.
Thanks.
And about the other problems?
Marcoa

[Suncuss]

that should help i think because there are two steps of the detection process. the first step is to just run the audio through the main BirdNET model, that will give you a list of probability of all species. Sometimes, multiple species will all ranked at very high confidence, maybe just becuase the bird calls are pretty similar. Then at the next step, the location and the week of the year info is feed to another model (the meta model), that model take's location and time info and predict what is the probability of seeing a bird in that particular location at that time of the year. This would help to further narrow down the list.

For example, it might detected some bird from your audio stream, with high confidence, but then the meta model says the chance of that particular bird appears in north pole in that time of the year is low, so it will reject that detection.

So fixing the location shold help with detection as well.

(Location filter is that know that controls it) at default, its behavior is that the mete model output has to be at least 3% for it to not be filtered out.

[jimenoma]

HI Suncuss:
The station's performance is now similar to that of the previous Nachtzuster software. The number of species identified hovers around 25–30 per day.
However, the absolute number has increased significantly. In fact, on some days, it ranks first globally, thanks to the bee-eater colony located very close by.
Thanks and regards,
MarcoA

[Suncuss]

That is great to hear! Jeje. One thing that might explain the differences is that default Confidence Threshold in BirdNET-PiPy is 0.6 but in Nachtzuster it might be 0.7. So you could try increase that to 0.7.