24/7 Raspberry Pi/birdNET monitoring (1 Viewer)

[Jon.Bryant]

Rechargeable LiOn

A Bing Search led me to this quote in a refence.

'How Does Cold Weather Affect Lithium-Ion Batteries? ​

Have you ever had to abandon an outdoor project or a road trip because your lithium-ion batteries had died prematurely due to the cold weather? Well, cold weather is hard on lithium-ion batteries and can significantly reduce their efficiency and performance, regardless of their reputation as one of the best batteries in cold weather.

Lithium batteries discharge an electric current when the transfer of lithium-ion occurs from the graphite anode (negative electrode) to the cathode (positive electrode). This process slows down in cold weather thus weakening their power. As the temperature drops, the lithium ions will just coat the anode (lithium plating) thus increasing the resistance of the electrolyte and making fewer lithium ions available to cause the flow of electricity. This can reduce 20-30% of the rated battery capacity, although ideally, lithium-ion batteries should operate at 98-95% of the rated capacity.'

Although many batteries say that they work down to -20 C. From the research I have done I think the Ultimate Lithium disposable batteries are they only ones that continue to work at a high percentage of rated capacity.

 

[Jon.Bryant]

I think a 30,000mah powerbank would cover my needs

I also found this on the web to substantiate the cold weather performance of Energizer Ultimate Lithium (https://data.energizer.com/pdfs/l91.pdf).

Trying to backtrack from the battery life calculation, I think the drain on the PUC is about 160mA (I could be wrong),. At that level there should be no real dip in performance, even in the worst of a UK winter.

If you stick with a Lithium ion power bank, the other thing I read, is that you should allow Lithium Ion to warm up before recharging, as charging when cold can lead to 'cracking' and battery damage. You may therefore need a couple of power banks, so that you can run one while allowing the other to warm up in the house before charging.

 

[Jon.Bryant]

Stumbled across this article (Evaluation of BirdNet), when looking for data on BirdNet (and PUC) data consumption.

Interesting review of the accuracy of the system, but I thought it was particularly interesting that ‘Similarly, BirdNET and Kaleidoscope Pro were able to detect 76% and 78% of the vocalizations detected by a human observer, respectively’. This suggests that unfortunately there is no substitute for learning vocalisations and plain old listening.

Of course BirdNet and the PUC do an amazing job, when there is no one there to listen. 76% is infinitely better than 0%. And if you don’t know the vocalisations, again the BirdNet detections are better than total bewilderment - but of course the accuracy data suggests a need for human verification.

A little disappointing that the data and specification of the PUC is so sketchy - do the mics outperform a typical phone for example (perhaps increasing accuracy and distance for detections). A study to show if the PUC outperforms the app simply run on a mobile, would be a start. Typical data consumption would be important to me, as I can imagine using the device on holidays, where travel plans have more limited allowances.

I suspect that the design of the PUC is based on the concept of a static listening hub with WiFi connection - say a backyard (or perhaps an observatory or ecolodge garden). For me, to qualify as a mobile field device, it has to clearly outperform a mobile phone running the app - although GPS tagging is a nice feature.

 

[LittleBitOfBreadNoCheese]

Stumbled across this article (Evaluation of BirdNet), when looking for data on BirdNet (and PUC) data consumption.

Interesting review of the accuracy of the system, but I thought it was particularly interesting that ‘Similarly, BirdNET and Kaleidoscope Pro were able to detect 76% and 78% of the vocalizations detected by a human observer, respectively’. This suggests that unfortunately there is no substitute for learning vocalisations and plain old listening.

Of course BirdNet and the PUC do an amazing job, when there is no one there to listen. 76% is infinitely better than 0%. And if you don’t know the vocalisations, again the BirdNet detections are better than total bewilderment - but of course the accuracy data suggests a need for human verification.

A little disappointing that the data and specification of the PUC is so sketchy - do the mics outperform a typical phone for example (perhaps increasing accuracy and distance for detections). A study to show if the PUC outperforms the app simply run on a mobile, would be a start. Typical data consumption would be important to me, as I can imagine using the device on holidays, where travel plans have more limited allowances.

I suspect that the design of the PUC is based on the concept of a static listening hub with WiFi connection - say a backyard (or perhaps an observatory or ecolodge garden). Fir me, to qualify as a mobile field device, it has to clearly outperform a mobile phone running the app - although GPS tagging is a nice feature.

The test was quite restricted - the human observer wasn't blind to the species (Bittern only) and his detections were taken as the benchmark. It scored 59/63 measured at the detection window level. But the best strength isn't necessarily accuracy; bN doesn't rest or get bored or lose concentration.
Bittern seems like an odd choice of species from a birder pov although maybe makes more sense if you are a conservationist.

 

[Jon.Bryant]

The test was quite restricted - the human observer wasn't blind to the species (Bittern only)

I couldn’t work out whether this sentence in the report continued the sentence about Bitterns. As you say this is an odd species to run stats on - I would have thought that vocalisations carry so far and population density is so low, that ‘standard’ survey would suffice.

Nevertheless the article states that the
‘Recall rate has been estimated for just 19 North American bird species, with an average value of 0.43 ± 0.25 (range 0.09–0.90, 95% confidence interval (CI) 0.32–0.55, see Appendix S2). Recall rate can be considered low, as it is around half the average precision estimated in most studies (Table 2).’ As recall rate is the number of detections compared to the total number of actual vocalisations, 0.43 +- 0.25 still seems quite low.

 

[LittleBitOfBreadNoCheese]

Yeah they reviewed quite a lot of studies so I just read the Bittern one as it was the source of the 0.76 figure you quoted.

https://www.sciencedirect.com/science/article/pii/S1574954122003600?pes=vor

"BirdNET detected the species in 59 of the 63 (93.7%) recordings with known presence of the species, while Kaleidoscope detected the bittern in 62 recordings (98.4%). At the vocalization level, BirdNet and Kaleidoscope Pro were able to detect between 76 and 78%, respectively, of the vocalizations detected by a human observer."

I would still quibble about a certain amount of bias in these determinations. eg how many Bittern calls are there really in the data set... ;-)

 

[pppppvvvv]

An interesting thread, which can potentially give some new life to my audiomoths which currently gather dust.
I've ran audiomoth output files thru the birdnet analyser and results seem fine.
Massive time saving compared with scrolling thru audacity sonograms!

One additional way to eliminate false hits from a static deployment would be to disregard one-off calls, but pay a bit more attention if the species is repeating regularly in the file.

There are a few realistic/optimistic rare target species in my area such as spotted crake, Long eared owl, Quail and Stone Curlew, that this could be put to good use for.

 

[Jon.Bryant]

I've ran audiomoth output files thru the birdnet analyser and results seem fine.
Massive time saving compared with scrolling thru audacity sonograms!

Hi Peter,

I feel like I am missing something, can you advise how you did this?

I tried uploading long recordings through the BirdNet web portal, but I recall that it was rather cumbersome. Long upload times (despite fast broadband), and then only a single list of species that varied as I played the file. I don’t think that it provided a full list of detections with time offsets and allowed me to jump to the vocalisation. Did you run it through the web portal, or are you running analysis on your own device?

Kaleidoscope extracts snippets of the file for review (or at least provided a view of the relevant part of the file), but I found the free view hard to setup and the subscription version is too expensive for me. I was setting up the detection algorithm on loudness, length of sound, gap between sounds etc, rather than detection of a particular species. In theory the free version of Kaleidoscope should fast track review of NocMig recordings, by isolating sounds of interest, thus eliminating the need to visually scan hours of ambient noise sonograms. Potentially, I think you may ‘miss’ less with Kaleidoscope, but the downside is that you need to be able to ID what is detected.

 

[Jon.Bryant]

I've ran audiomoth output files thru the birdnet analyser and results seem fine.
Massive time saving compared with scrolling thru audacity sonograms!

Hi Peter,

I feel like I am missing something, can you advise how you did this?

I tried uploading long recordings through the BirdNet web portal, but I recall that it was rather cumbersome. Long upload times (despite fast broadband), and then only a single list of species that varied as I played the file. I don’t think that it provided a full list of detections with time offsets and allowed me to jump to the vocalisation. Did you run it through the web portal, or are you running analysis on your own device?

Kaleidoscope extracts snippets of the file for review (or at least provided a view of the relevant part of the file), but I found the free view hard to setup and the subscription version is too expensive for me. I was setting up the detection algorithm on loudness, length of sound, gap between sounds etc, rather than detection of a particular species. In theory the free version of Kaleidoscope should fast track review of NocMig recordings, by isolating sounds of interest, thus eliminating the need to visually scan hours of ambient noise sonograms. Potentially, I think you may ‘miss’ less with Kaleidoscope, but the downside is that you need to be able to ID what is detected

 

[pppppvvvv]

Hi, I've used the desktop version of birdnet-analyzer installed on windows laptop.

GitHub - kahst/BirdNET-Analyzer: BirdNET analyzer for scientific audio data processing.

BirdNET analyzer for scientific audio data processing. - kahst/BirdNET-Analyzer

github.com

The GUI application seems easy enough to use, and with a bit of tweaking the confidence threshold, and location coordinates, I was getting good results.

There is a minor problem in that python 3.12 is not compatible with tensorflow. This doesn't mean a great deal to me! But there is a command to check what species are being searched for based on the location, and this doesn't work for me. I should wipe it and start again with an earlier version of python.

 

[pppppvvvv]

Hi, I've used the desktop version of birdnet-analyzer installed on windows laptop.

GitHub - kahst/BirdNET-Analyzer: BirdNET analyzer for scientific audio data processing.

BirdNET analyzer for scientific audio data processing. - kahst/BirdNET-Analyzer

github.com

The GUI application seems easy enough to use, and with a bit of tweaking the confidence threshold, and location coordinates, I was getting good results.

There is a minor problem in that python 3.12 is not compatible with tensorflow. This doesn't mean a great deal to me! But there is a command to check what species are being searched for based on the location, and this doesn't work for me. I should wipe it and start again with an earlier version of python.

This isn't LIVE, I should mention, but collecting the files and running them after the event. The audiomoth file I was using was from last May.

Typical results output..

 

[Jon.Bryant]

I should wipe it and start again with an earlier version of python.

from some Python programming work I did, I am pretty sure you can have multiple versions of Python installed at the same time - I think last time I checked I have two versions. Of course you need to correct choose the environment you run the programme under.

 

[pppppvvvv]

from some Python programming work I did, I am pretty sure you can have multiple versions of Python installed at the same time - I think last time I checked I have two versions. Of course you need to correct choose the environment you run the programme under.

that would be better I think. I'm reluctant to start again, as it is working quite well.

 

[GreenRooky]

They have just created a forum for the PUC and birdweather community
BirdWeather Community

 

[Buckster]

I will write more when I have the chance - but my BirdNET Pi has completely revolutionised my view of the birds in my garden within weeks

I use an overclocked Pi4b (4gb version) with a USB mic, and attached to a 20,000mAH battery I think I get about 18 hours of use/recording

an example of how has changed my view - I have heard a Tawny Owl once since I moved to this house 4 years ago or so - I thought was a one off as have never heard since

within 2 weeks of putting BirdNET Pi out I've detected Tawny Owls overnight across at least 3 nights and I haven't put the Pi out every night. It has also detected multiple flying Water Rails, Moorhens, Mallards, Wigeon over night too

2 Tawny Owl recordings from last night:

around midnight a male Tawny Owl calling

around 2am a female Tawny Owl calling twice followed by a male replying

 

[LittleBitOfBreadNoCheese]

PUCs continue to spread out, just found one near Chernobyl

Chornobyl WildLife :: BirdWeather

app.birdweather.com

 

[LittleBitOfBreadNoCheese]

Hi, has anyone got any hands on experience with the birdweather recording device? BirdWeather - a living library of bird vocalizations.

I've previously used a voice recorder, then an audiomoth and then a parabolic dish for nocmig recording (with the dish being best by far). I've given up on Nocmig a couple of years ago because the post-processing took too much time.

The above device looks good on paper to solve this, however concerned about both sensitivity and single call id (would this be logged as unknown or completely ignored). Similarly if no matching record in library what happens with the call.

Hope you don't mind me moving your question into the main thread.
I'm not sure what your first question means, perhaps you can clarify.
Non-IDs get deleted. The server does hold a handful of non-bird sounds like car engines, barking dogs etc. Human voices are screened out.

 

[GMS]

Hope you don't mind me moving your question into the main thread.
I'm not sure what your first question means, perhaps you can clarify.
Non-IDs get deleted. The server does hold a handful of non-bird sounds like car engines, barking dogs etc. Human voices are screened out.

Thank you. What I'm trying to understand is of there are active users of the birdweather device and what is their experience compared to other options.

Deletion of non-IDs is a big drawback for me. Anyone who's done nocmig knows that unknown are a regular feature. Also it really depends on the library.

 

[LittleBitOfBreadNoCheese]

Thank you. What I'm trying to understand is of there are active users of the birdweather device and what is their experience compared to other options.

Deletion of non-IDs is a big drawback for me. Anyone who's done nocmig knows that unknown are a regular feature. Also it really depends on the library.

Bear in mind that BirdWeather is a client-server system although it can be used in the field in which case it will upload all the files and ID them when it reconnects with your wifi. If you want to manually grab them too (for separate analysis) go ahead. There is a 'purge after upload' setting but this pertains to field use.

But most people are using them adjacent to their house over wifi and they go straight to the server. And we are talking potentially thousands of detections a day. You can play around with the settings to maximise the chances of some kind of ID taking place. ID soundscapes can be replayed off the server.

Compared with continuous audio recording to some giant .wav file it is a very different concept.

 

[LittleBitOfBreadNoCheese]

Top 3 stations yesterday had over 10000 detections to upload.