Many times hydrophone recording contains unwanted noise like 'wave noise'. How to deal with this noise at the source? Is there any way to deal with this in the analysis stage? as abrupt noise impacts the detection of targeted sound.
There are a couple of principles to enhance the signal to noise ratio
- bandpass filtering removes out-of-band noise
- median filters attenuate impulsive noise, when one is interested in tonals.
- matched filter is signal has characteristic waveforms
- notch filter if there is a particular frequency interfering.
Edit: As Rasmus indicates, better avoid generation of noise in the first place and design the system such that it does not generate so much noise. OK, it is not always easy. E.G. a chain is a clever way to add weight and to minimize vertical impulsive stress from recorder. But, system design was not mentioned in OP.
$\begingroup$ If the signal you're looking for has its energy concentrated in a few frequencies, broad-spectrum noise can be dealt with by dropping any frequency bins below a "background noise" threshold for that bin, i.e. a noise-gate in each frequency bin after an FFT. (Audio noise reduction: how does audacity compare to other options?). Other modern techniques include Non-Local Means and Recurrent Neural Networks, which FFmpeg has filters for. No idea if they'd be appropriate for analysis, rather than human-perceptual. $\endgroup$ Jul 27, 2022 at 18:09
$\begingroup$ @Peter Cordes: Is that a comment to my answer or to the OP? $\endgroup$ Jul 27, 2022 at 18:23
$\begingroup$ A suggested addition to your answer, because I don't know which of those noise-filters are appropriate for bioacoustics. I've only ever worked with audio recordings in air, of speech / music intended for human listening, not scientific analysis. If you think my comment's useful as a separate answer, I can post it as such. $\endgroup$ Jul 27, 2022 at 18:25
$\begingroup$ I guess, you are describing some sort of bandpass filter (equivalently suppressing unwanted frequencies), right? $\endgroup$ Jul 27, 2022 at 18:30
$\begingroup$ A noise gate for each FFT bin is like an adaptive band-pass I guess. It zeros out every frequency except ones that have significant energy above the noise floor. The basic model is to do an FFT and threshold each bin, setting it to zero if below the noise threshold. (Then convert back from frequency domain to time domain to filtered PCM audio samples.) Choosing the right threshold for each bin (over time) is the trick, e.g. based on a profile of background noise if you have one. It works well for white noise "hiss", not for clanks or hums. $\endgroup$ Jul 27, 2022 at 18:51
You originally asked for 3 different noise sources, and you ask how to deal with it at the source (and later in analyses):
- Bumping on sea floor: If your hydrophone is bumping directly on the seafloor (unlikely given your fine illustration) then that's the real problem, and hard to solve for you without more information on the setup. If the bumping is the noise from your anchor/acoustic release bumping on the seafloor, then either have your top float sit lower in the water column (less wave heave effect) or make your anchor softer (wrap it in something soft and bio-friendly).
- Chain noise: Use heavily bio-fouled chains (leave them out in the water for a long time before deployment), or wrap in other bio-friendly soft material. Alternatively, don't use a chain...
- Wave noise (the currently mentioned issue): This is presumably from the float? is so, try to deploy it deeper. If this is surface wave noise, there's not much you can do at the source.
In analyses: As @WMXZ writes there are many options for DSP solutions, but without knowing more about what sounds your are interested in keeping, it's hard to suggest a good approach. You cannot just eliminate all wave noise in the analysis stage, but you can likely make filters that will improve the detection/quality of the signals you're looking for. Please add some information as to what kind of sound you're interested in detecting.
$\begingroup$ OP modified question to separate the issues (see his other Q), if you are able to update your answer. Thanks $\endgroup$– ShannonJul 29, 2022 at 2:30
Not sure if this is something you could consider, but my lab just had these floats custom made to hold a SoundTrap and F-POD with the goal of stabilizing the hydrophones, reducing strumming, etc. They'll be deployed at the beginning of August, so I’ll let you know how it goes!
$\begingroup$ Whoa - I've never seen something like this before. How exactly will it be deployed? $\endgroup$– ASimonisJul 27, 2022 at 21:00
1$\begingroup$ I don't know all the deployment details (I'm probably won’t involved with these particular cruises) but there will be an EdgeTech acoustic release below of and chain linked weights below that -- similar to the schematic above. It's being deployed off Oregon. $\endgroup$ Jul 27, 2022 at 22:36
$\begingroup$ Wow. This is something awesome, and I didn't know such a thing exist. Thank you so much for informing me about this equipment. $\endgroup$ Jul 28, 2022 at 4:04
2$\begingroup$ Having the hydrophones so close to the floating material worries me a little bit. (air filled) floating materials, like this foam as it seems, will shield sound from below and reflect sound from above. $\endgroup$ Jul 28, 2022 at 5:19
$\begingroup$ Good point. I'll chat with the crew that made them and see what they think about that. $\endgroup$ Jul 29, 2022 at 15:21
If you can deploy multiple hydrophones (either on same cable or on separate cables) and synchronize their recordings, it is possible to filter signals based on their arrival angle.
Two hydrophones vertically separated on same cable can distinguish arrival angle in the vertical direction, allowing you to focus on signals coming from the target depth horizontally.
$\begingroup$ Could you give a link to a paper describing or using such a method please? or from another stackexchange thread? $\endgroup$– NoilSep 24, 2022 at 9:38
1$\begingroup$ @Noil The usual term is beamforming. Here are some resources that appear relevant, though I don't know very much of the subject: Moura et al 2009, Kong et al 2021. And e.g. Turbulent Research recorders have synchronization support. $\endgroup$– jpaSep 24, 2022 at 16:44
If nothing can be done on instrument design, and you have a lot of data with consistent noise inputs, machine learning methods are very capable of learning the noise sources in your data and excluding them from further analysis.
Several audio mixing programs (DAW = Digital Audio Workstation) has noise reductions functions. It is possible that if you use one of them you could get a good reduction. Some of them work on taking a no-signal noise signature first. The algorithms seems to be proprietary. A free example with source code is available in the program Audacity. To say one single product (expensive though) used in recording the Izotope Rx comes to mind. There seems to be specialized tools used in forensic evidence situations, I have no names there though.