CityNet seems to be overestimating anthropogenic noise

[KirstenMP]

I'm using CityNet to estimate the levels of anthropogenic noise at sites in Switzerland and France where I've been recording birds and frogs for a freshwater-soundscape project. I have the code working, but the algorithm is overestimating human-generated noise, with the blue 'anthrop' line on the plots sitting at or near 1 for many recordings where I judge there to be very little anthropogenic noise.

I've attached an example here, with the .wav file.

Any suggestions welcome - thanks! 🦖

2MM05468_20250529_190000 4.wav

[mdfirman]

Thanks for your interest in this project Kirsten!

And sorry this model doesn't seem to be working well for you.

I've got a couple of thoughts here. I will caveat this though by saying that I don't actively maintain or improve this model at all now.

1. Our model was trained and evaluated on sounds from places around London with a specific type of microphone. So extrapolating to other locations and recording setups is not guaranteed!
2. There are a lot of sounds all at once in your recording. While they all sound biotic, I wonder if the sheer number of sounds is causing the model to think there is a lot of anthropological sound. (Has the model learned that "more noise means more likely to be human noise"?). No suggestions on this point, just an observation.
3. In particular, when I boost the volume it sounds like there is a low-frequency background hum throughout almost all the recording. I wonder if you were able to remove that (e.g. with a high-pass filter, which should remove low-frequency noise and keep high frequencies) you might see the anthrop level reduce? In our training set, a lot of the anthrop sound was aeroplanes and cars, and I wonder if that low-frequency hum is giving you false positives here.

Overall though this project is now getting on for 10 years old and I think there must be better options out there for you? Do please keep me updated with how you get on!

Thanks– Michael

[KirstenMP]

Thanks Michael,

My suspicion is that the acoustic environment of London is very different to that of my survey sites, leading to a poor match between the training data and the sound files I'm running. I will try removing the background hum and see if this improves the estimation of anthropogenic-noise levels. :-)