Is a limit of detection published for the AudioMoth? This number was really important in instrumental analysis. I believe it is defined as three times the standard deviation of ambient measurement above the mean.
I think someone made a detailed study of this within the last year (after all of the above comments), and that it might be available on GitHub or somewhere similar - does anyone recall the citation I'm referencing? I can't refind it!
I am not sure that Limit of Detection is an appropriate measure when trying to detect impulsive signals against a noise background. Detectability is usually considered to be based on S/N ratio but even that is a tricky measure in acoustics. Noise level could be the broad-band noise level across the full recording spectrum - which is relevant if trying to detect signals in the time domain - e.g for a zero-crossing detector, but detectability in a sonagram is different because you are looking at signal against a noise level within an FFT frequency band and so consideration of noise/sqrt(Hz) would be a better measure. But then for structured signals such as bat calls detection is not limited to a single frequency-bin but is distributed across the bandwidth of the signal. Couple all of that with the fact of having an uncalibrated microphone with a sensitivity that varies with frequency and it is clear that there is no simple measure which can define a limit of detection when applied to recording animals.
Said another way: If I see a value in my spectrogram of -250 dB is that a number which could actually be recorded by an AudioMoth (at a given frequency).
@Barry Moore To give a definitive answer I would need to know a bit more about how you are using the AudioMoth. A measurement of the broadband noise level will give some indication, but that will also depend on the sampling rate (and therefore effective bandwidth) and the AudioMoth gain setting. Noise level will not strictly be proportion to the Nyquist frequency bandwidth because higher frequencies in the microphone and system will be aliased back into the bandwidth of interest. In practical terms, at 192ksps and lo gain, my background noise level (In Audacity waveform, log scale view) is around -50dB which is effectively the dynamic range of the recording. In a power spectrum averaged over half a second, the noise floor drops from -72dB at 1.5kHz to -82dB at 96kHz, but within that there are bands that are noisier, -67dB at 17kHz, -71dB at 36kHz and -75dB at 60kHz with a 'whistle' artefact of -73dB at 75kHz. (Reducing the input gain in the AudioMoth - a hardware mod, should give an increase in dynamic range but I have yet to try it out). On top of all that is the problem of clutter as opposed to noise - i.e. audio signal which is not from your desired target. Walking through long grass produces lots of noise, equipment goes chink, all sorts. Detecting genuine signal is then much more a problem of pattern recognition in a 2-dimensional image (the sonagram) and that is a whole different ballpark. On the other hand sensitivity - i.e. maximum detection range at which you recognise a 'pattern' is also a function of frequency being affected by the microphone frequency response (see the Knowles data sheet) and atmospheric absorption. Basically trying to put numbers to all of these is not wonderfully helpful, but in general the AUdioMoth is comparable to other recorders on the market apart from the more limited dynamic range, and for identification purposes overloading is not a severe problem, although it matters if you want to study signal structure.
Barbastelle pass recorded with AudioMoth 192ksps Lo Gain
By way of illustration, the attached image is an Audacity sonagram (1024FFT, Hanning window) of the start of a Barbastelle pass. The first pulse (highlighted) is close to the limits of detectability, but would be hard to identify from the power spectrum and you could not say that anything in the power spectrum above a certain level was signal rather than noise or clutter. In this case the pulse is just identifiable on the spectrogram, aided in this case by the context of the rest of the pass. Spectrogram visibility also depends on such things as monitor resolution, FFT window size, and very much on the time axis. Zooming out would make this pulse disappear completely, and zooming in might make an even earlier pulse visible, but of course the more you zoom in the longer it takes to analyse a recording, and ultimately it all depends on what information you are trying to extract or quantify from the analysis.
I think someone made a detailed study of this within the last year (after all of the above comments), and that it might be available on GitHub or somewhere similar - does anyone recall the citation I'm referencing? I can't refind it!
Buy best USA made cat5e plenum, cat6 plenum Cables, Fiber optic Cables ,Coaxial cables and network management tools. http://atscables.com/
I am not sure that Limit of Detection is an appropriate measure when trying to detect impulsive signals against a noise background. Detectability is usually considered to be based on S/N ratio but even that is a tricky measure in acoustics. Noise level could be the broad-band noise level across the full recording spectrum - which is relevant if trying to detect signals in the time domain - e.g for a zero-crossing detector, but detectability in a sonagram is different because you are looking at signal against a noise level within an FFT frequency band and so consideration of noise/sqrt(Hz) would be a better measure. But then for structured signals such as bat calls detection is not limited to a single frequency-bin but is distributed across the bandwidth of the signal. Couple all of that with the fact of having an uncalibrated microphone with a sensitivity that varies with frequency and it is clear that there is no simple measure which can define a limit of detection when applied to recording animals.
No, I haven't come across that measure. I'll look into it.