To test this feature, visit your live site.

Bird species ID with background noise using R (warbleR, monitoR)

Hello

I’m using AudioMoth devices (version 1.0.0) to record bird sounds in African savanna with settings of: 1-min recording length; 32-Htz sample rate (SR); “medium” gain, using; a “SanDisk Extreme” 32GB micro-SIM card with 1.5V Energizer Alkaline Batteries.

I’m trying to detect bird songs using the R packages warbleR (https://rpubs.com/marcelo-araya-salas/110155) &/or monitoR (https://rdrr.io/cran/bioacoustics/f/vignettes/tutorial.Rmd) using the following workflow. 1) identify “event windows” (i.e. a range of sound-attributes associated with the bird-call of interest, that may also include other similar bird calls), 2) extract 22 – 27 specific sound-attributes included in the event windows and label as bird of interest (yes) or not (no), 3) predict bird-calls of interest in other sound recordings using a Random Forest Model. I can use the workflow to make good predictions using bird-calls imported from Xeno-canto (https://www.xeno-canto.org/; 44 Hz SR), however I can’t use the same protocol to make similar predictions using my AudioMoth sound recordings.

From the sonograms (see attached; continuous time on x-axis, frequency on y-axis in Hz), it seems that the AudioMoth recordings have a poorer resolution and higher background noise than the Xeno-canto ones (even when the 44-Hz Xeno-canto calls are downscaled to 32-Hz). Could this causing the problem? Alternatively, perhaps there is an issue with using the AudioMoth .WAV files instead of .wav files in these packages?

Has anyone conducted similar analyses and experienced similar problems? Any help would be greatly appreciated. I am happy to share specific sound-clips and script if needed.

Thank you for your help!

9 comments

9 Comments

sarahd

Jun 08, 2020

Hi everyone,

I'm hoping someone can help me with a similar issue I'm having with warbleR as gordonc--I'm trying to figure out the appropriate parameters for my autodetect function. I have adjusted multiple parameters a bunch of times to try and have the program "find" the song I'm interested in, but I'm not having any luck. The .wav files that are being used here are all from xeno-canto, so they all definitely have the song in the file that I'm interested in. It will pick bits and pieces of the spectrogram but it won't read the entire thing. I've attached a .tiff file and .wav files for you all to see as an example, and I've included the code as well.

Thanks so much in advance!!

Sarah

sub <- wavs[sample(1:length(wavs), 3)]

sub

autodetec(flist = sub, bp = c(1, 8), threshold = 15,

          mindur = 0.05, maxdur = 3, envt = "abs",

          ssmooth = 600, ls = TRUE, res = 100, flim= c(0, 8),

          wl = 2000, set =TRUE, sxrow = 10, rows = 15, redo = TRUE,

          it = "tiff", img = TRUE, smadj = "end")

Robert Manzano

Apr 14, 2020

Hi gordonc,

We will also be working with Audiomoth+warbleR package monitoring bird migration. How long does the autodetect takes to run through a night's recording (10-12h)? Ours need 5h to do so, which I it's a lot.

Many thanks !

Rob

gordonc

Apr 15, 2020

Replying to

Hi Rob

Good to hear that you’re also trying to analyse bird data using warbleR. It was taking me 5-6h to process ~ 3300 1min audio clips. I was using four parallel cores (using the “parallel” call to do this, say in the “autodect” function) on my computer, however I imagine you could quicken the processing time considerably if you have access to a larger processing computer by specifying more cores. I imagine that processing speed also relates to the size of your files, so this is a consideration too.

Out of interest, how are you processing your bird calls? My aim is to build a model to predict specific bird calls within my sound clips (i.e. present / absent), using data from Xeno-Canto and my own recordings as training data. To date, I have been able to extract the relevant information using warbleR and then build a good Random Forest model which predicts well on the training data. My problem is that this model has poor predicative accuracy on my testing data; that is, the 3300 audio clips I want to predict calls on. I assume the problem relates to the fact that I am not including enough different types of “no” calls in my training data, which would explain why my model is performing well on my training, but not testing, data.

Anyway, hope this helps

Cheers

Chris G

Robert Manzano

Apr 16, 2020

Replying to

@gordonc

Hi Chris,

Many thanks for your quick reply indeed. How do you set the parallel argument to use 4 cores? I've tried setting "parallel=2" (I'm using a i5 with two cores) with no difference in performance. Perhaps there is a different way to set the argument?

We will be doing the ID manually as we don't have enough statistical knowledge (and time to learn this :( ) to set an autonomous ID process, but I've seen papers talking about machine learning processes including neural networks. Take a look at the following paper, which I understand it's a good review of the topic:

https://onlinelibrary.wiley.com/doi/full/10.1111/jav.01447

I bet you have already consider it, but does your testing data have similar SNR ratio to your training data? I know this is a critical point when doing automatic bird call detection and ID. I understand you are also using R for the random forests.

Are you focusing in only one type of bird call (e.g. one species)?

Cheers,

Rob

gordonc

Apr 17, 2020

Replying to

@Robert Manzano

Interesting. Yes, that was how I specified parallel cores in my script, which seemed to work well. Perhaps try bumping up the number of cores and see what happens.

Thanks for the advice re SNR. I have considered this, however perhaps re-visiting it in a more detail will help.

Cheers and good luck.

Chris G

Jan Borst

Sep 18, 2019

Nt! Good to hear you're making progress. Now you can label your events and then start "training" your Random Forest model with the variables of the autodetection. Therefore you first have to find out which variables will give the best results (this will take some trial and error to find the best results).

Good luck!

gordonc

Sep 18, 2019

Thank you for the advice Jan. Given your help I've now been able to find event-windows in my AudioMoth data (i.e. step one of my workflow). My problem was that I was focusing on the Freq and Duration attributes in autodetect and not so much on the Amplitude attributes. I found that I was able to reliably detect my event-window by fine-tuning the "threshold", "ssmooth" and "wl" attributes (see attached).

Again, thank you for your help

Jan Borst

Sep 18, 2019

Hello, Do you know where in the progress the problem starts? If it starts at step 1 of your workflow (finding the audio effents), you can try to change the sensitivity of the autodetection. Notice that in the tutorials of the packages they only use A-quality recordings from Xeno-Canto. So the parameters they use in the tutorials will work fine for recordings with similar quality. If you have recordings with more noise or calls that are not as loud, the autodetection will not detect the calls or only a small part of a call (making classification a lot more difficult). I'm using the bioacoustics package (threshold detection) in R to detect bat calls on my Audiomoth-recordings. Because bat calls are usually above the noise level this works fine. I've tried the same script (with different parameters) on some bird recordings that I've made with a Audiomoth. Because bird sounds are much lower and therefore more disturbed by background noise it was a lot harder to find the right parameters for the autodetection. I have not tried to classify this bird calls yet, but only finding them with the autodetection wasn't that easy. So the first step to get some result is to fine-tune the autodetection: run the autodetection with different parameters and compare the results in the data tables with the spectrograms of the recording to find the best parameters for the detection.

Hope this will help. Greetings, Jan