Processing time may be lengthy when the system has to generate spectral data on-the-fly. Please limit search request time ranges to 3 months or less for: PDF spectrograms, PNG spectrograms and MAT spectral data files with non-default data product options. |
This data may be diverted or delayed for security reasons. Read on below for more information. |
ONC hosts a large number of hydrophones (169 over 11 different types and growing). This instruments produce audio data with sounds at a wide range of frequencies, having applications in seismology, marine mammal studies, ship noise and more. Hydrophone spectral data (PNG/PDF image files of the spectrogram, FFT and MAT data files) are provided as a summary of the audio recording and for detailed analysis. With spectrogram images and data, users can determine the sources and nature of sounds they can hear in the audio data files: passing ships make hyperbolic curves, whales make chirps and resonances, etc. Spectrogram images, particularly the PDF format, are a great way to scan through a vast amount of data quickly, looking for events. Listening to this data would be very time consuming, plus the human ear is not capable of picking up some high or low frequency sounds that may be quite interesting.
Spectrogram image files (PDF/PNG) are available for all hydrophones and hydrophone arrays. For hydrophones located on low-bandwidth observatories (Cambridge Bay, Brentwood Bay, etc), audio data files may not be available. In this case, audio data is stored on site to be retrieved periodically during site visits and uploaded to the archiving system. To provide live data, low-bandwidth observatories upload spectrogram FFT files as they are much smaller than the full audio data. FFT files may be used to generate and fill in spectrogram plots when full audio data is not available (see the Spectrogram Source data product option option). MAT files store pre-generated one-minute ensemble-averaged spectral data, primarily for use in spectral probability density plots, and are available for download directly and quickly, while the full spectra are also available as an option on the MAT files (much slower to generate).
Oceans 3.0 API filter: dataProductCode=HSD
Given the sensitive nature of hydrophone data, the military has authority to completely divert the data and/or filter it as required. When filtering occurs, the file-name is appended with '-HPF' or '-LPF', corresponding to high and low pass filtered data (common prior to 2016, rare after this time). Often, the military diverted data will be reviewed and returned at a later date with or without filtering. See here for more information on the diversion of hydrophone and seismometer data. Data product options are provided to sort out the various types of diverted data (see below). Not all options are available for all devices: for instance, in some cases, we may not offer '-LPF' spectrograms even if there are '-LPF' audio files because such a spectrogram would offer so very little useful data. The option sets are also responsive to location (some locations are not subject to diversion). For icListen HF hydrophones, FFT spectral data are often available during a diversion and will be used to produce spectrograms to fill in data gaps until audio data is returned.
Spectrograms may take some time to generate. PNG spectrograms files are normally pre-generated and stored, so their retrieval is quick. However, PDF spectrograms are not pre-generated, but will be computed on-demand from stored source WAV or HYD files. Processing speed is nominally about 5 to 20 seconds per 5 minute audio file. The final step of PDF generation, aggregating the individual spectrograms, takes more time for a larger number of files. In all, requesting a day's worth of PDF spectrograms may 20 to 60 minutes. Please be patient, a day's worth of data is roughly 16 GB of WAV files, 500 MB of MP3s, 160 MB of PDFs, 200 MB of PNGs, 100 MB of MATs. A single hydrophone can generate terabytes of data in a year. It is a vast amount of data to store and serve to our users.
Spectrograms, spectral data and spectral probability density are calibrated to absolute sound pressure level whenever the calibration is available. Calibration data can be found in the device attributes page for each device, e.g. https://data.oceannetworks.ca/DeviceListing?DeviceId=23159. Calibrations are time sensitive and frequency specific. For instance, if a 48 kHz hydrophone only has calibration for the first 1600 Hz, then only the calibrated frequency bins will be shown, even though the device is captures signals up to 24 kHz. We will endeavour to calibrate as many hydrophones as possible: the icListen hydrophones in particular are more readily calibrated. When calibration is available, the units on the spectrograms will be dB re 1 uPa RMS and SpectData.isCalibrated in the mat files will be true or 1. When calibration is not available, the units on the spectrograms will be dB re full scale and SpectData.isCalibrated in the mat files will be false or 0. Note that spectrograms produced from FFT spectral data use a single point calibration and are not as accurate. The attributes HydrophoneSensitivityVectorPartXX contain the frequency of the leading edge of the hydrophone sensitivity bins and HydrophoneSensitivityVectorBinsLeadingEdgePartXX contains the Sensitivity vector for hydrophone calibration, where XX is the part number of an array containing these numbers (which are required to be split into parts due to limitations in the number of characters for the attributes in the database). A second set of device attributes with "Post" inserted in their name ("HydrophonePostSensitivityVectorPartXX") records the post-deployment calibrations, often carried out by ONC HydroCal. The post-deployment calibrations are not used in the spectral data calculations.
Spectral data is generated from the source WAV, FLAC or HYD files and makes use of the calibration data described above. The procedure outlined in Merchant et al. (2012), see the reference here: https://asa.scitation.org/doi/10.1121/1.4754429. Here is the procedure in code form: compiled snippets of our operational MATLAB code (which may change). Please note that the following code isn't directly runnable, it would need some modification; it is provided as documentation. Comments have been added for clarity.
%% read the source file (WAV or FLAC format
[H.data, H.sampleFrequency] = audioread(sourceFilePath); % don't use native format with wavread - want to scale to +/-1
Hinfo = audioinfo(sourceFilePath);
H.numBits = Hinfo.BitsPerSample;
%% set up the spectrogram parameters constrained by the calibration -> makes the time resolution variable, spectral resolution is set
% sensitivityBinFrequencies, sensitivity are read from the calibration file, decimationFactor is rarely used (usually value of 1) and badly named, it changes the resolution
j = 1; % channel index, set to one for this example
nFFT = H(j).sampleFrequency / median(diff(sensitivityBinFrequencies)) * decimationFactor;
windowLength = nFFT;
numPeriods = floor((length(H(j).data) - windowLength) / (1 - overlapFactor) / windowLength + 1);
timeResolution = numPeriods / (length(H(j).data) / H(j).sampleFrequency);
spectralResolution = H(j).sampleFrequency / nFFT;
%% calculate the spectra
% window the data with a Hanning window
% overlapFactor is 0.5 for spectrograms, but is 0 for spectral probability densities, including the one-minute averaged spectral MAT files
hannWindow = hann(windowLength, 'periodic');
windowedData = zeros(windowLength, numPeriods);
for k = 0:numPeriods-1
indexAdj = floor(k * (1-overlapFactor) * windowLength);
windowedData(:, k+1) = H(j).data((1 + indexAdj):(windowLength + indexAdj)) .* hannWindow * 2;
end
% to save memory, clear the data struct after windowing, won't need it again
H(j).data = [];
% calculate the FFT on the windowed data matrix (yes, it works on the correct dimension). Clear the windowData to save memory.
thisFFT = fft(windowedData, nFFT);
windowedData = [];
% calculate the single-sided FFT magnitude for the peak amplitude, scale PSD for Hanning noise power bandwidth. Clear the FFT to save memory.
psd = abs(thisFFT(1:floor(nFFT/2)+1,:)) * (2 / windowLength / sqrt(1.5));
thisFFT = [];
% calculate the time and freq axes for the PSD
sTimeSec = (1:numPeriods) * (windowLength * (1-overlapFactor))/H(j).sampleFrequency; % this accounts for the non-existence of the zeroth and nth windows
sFreq = (0:floor(nFFT/2)) * spectralResolution; % this accounts for the nFFT being an odd number
%% calibrate
% interpolate the calibration on the frequency bins - this is in case there's some discrepancy - normally doesn't affect anything
sensInterp = interp1(sensitivityBinFrequencies, sensitivity, sFreq);
% only use bins that are with in the range of calibration - in the case of LPF or HPF data exclude data outside of the cutoff frequencies
% the default values for lowPassCutoffFreq, highPassCutoffFreq is NaN, so goodFreqLI doesn't trim any frequencies
goodFreqLI = all([sFreq >= nanmax(-1, highPassCutoffFreq); sFreq <= nanmin(Inf, lowPassCutoffFreq); sFreq <= max(sensitivityBinFrequencies)], 1);
sensInterp = repmat(sensInterp(goodFreqLI).', 1, size(psd, 2));
psd = psd(goodFreqLI, :) * (1/sqrt(2)) * (2 ^(H(j).numBits - 1)); % scale peak amplitude to RMS, then convert +/-1 range data to native scale
psd = 20*log10(psd ./ 10.^(sensInterp/20)); % apply the calibration - sensitivities to correct spectral response
sFreq = sFreq(goodFreqLI); % update the frequencies
%% Compensate for Hanning window: find the max PSD for the windowing function and subtract it
hannFFT = fft(hannWindow, nFFT);
hannPSD = 2*abs( hannFFT(1:floor(nFFT/2)+1,:) / windowLength );
psd = psd - 20*log10( max(hannPSD(1:length(sFreq))) );
|
The end product of the above core code is the "psd" power spectral density, "sFreq" frequency bins, "sTimeSec" time bins, with the latter two matching the dimensions of the "psd" matrix. Future changes will include GPU processing and perhaps some refactoring (this code doesn't meet our code development standards as the standards are newer than the code!). After the spectral data is calculated a number of additional steps maybe applied prior to producing the spectrograms or MAT data files. This includes careful downsampling (in time and frequency) for spectrograms so that each pixel is a data point (so that the rendering doesn't distort the log scale data, usually the dimensions of the "psd" matrix are much larger than the dimensions of the images we render to screen) and we also do one-minute downsampling for the default one-minute spectral MAT files; all of the above downsampling is really just box-car style re-binning with log scale averaging. The raw calibrated spectra, one-minute averaged spectra and the data plotted in the spectrogram plots are all available via spectral MAT data files using "Spectral Data Downsampling" option (see the option section below).
{include: Hydrophone Data Acquisition and Diversion Mode}
Generally, this format is a spectrogram plot of 5 minutes of hydrophone/audio data. Here is an example PNG spectrogram taken from a hydrophone as Cascadia Basin:
Since spectrograms are stored for fast retrieval, users may see older versions (such as the example above), which have different titles, logos, etc., however the data is the same. The colour scale is fixed to facilitate comparisons between multiple spectrograms. Some spectrograms are calibrated with units of dB re 1 μPa. Non-calibrated spectrogram have a colour scale that is relative to the full range of the source audio file: the extreme values in the audio file are scaled from 0 to 1, so that the dB scale is from -120 dB (0.000001) to 0 dB (1). The spectrogram is generated by a modified Welch method: the data is windowed in time (Hann window 50% overlap) the length of each window is equal to the length of the FFT and the power spectra of each windowed segment is then a column in spectrogram data matrix, the rows are the different frequencies. If calibrated, the calibration range and resolution sets the length of the FFT, while if not calibrated, the length is set by optimizing the trade-off between temporal and spectral resolutions so that the spectrogram data matrix has an aspect ratio that's similar to that of the image file to be generated. Quite often, there are far more columns and rows in the spectrogram data matrix than there are pixels in the image file. In that case, the spectrogram data matrix is downsampled by linear scale box car averaging in both time and frequency to closely match the size of the image in pixels. This occurs after the spectrogram data is calculated, but prior to printing the data to the image file. Standard image renders would distort the logarithmic scale data with linear scale, anti-alias low-pass filtering or averaging, or if not, they would alias the data by decimating it. (There are actually two rendering stages: at image file creation and when the file is printed or displayed on screen, so it is best to view or print spectrograms with as many pixels as in the image file.) This downsampling only occurs for newer versions of the spectrogram data product. Newer versions also have a fixed relationship between time and position on the plot (1 pixel is about 0.3 seconds): if less than 5 minutes of data is provided in the audio source file, the spectrogram will not be stretched to fill the x-axis, but instead the x-axis will be shorter than usual. This will allow us to stitch together spectrograms in our data viewers (to be developed, prototype version exists). An exception to the fixed duration of spectrograms is when there is a varying duty cycle, i.e. where the duration and sample rate of the source WAV files vary; for example: 60 seconds at 128 kHz and 12 minutes at 16 kHz sample rate. The device attribute 'Spectrogram_ModeDurationDPO' is used to store the duty cycle parameters (duration and sample rate pairs) for use in data product generation. Currently, the deployments with varying duty cycle have only two sample rates: low and high. A data product option is offered for users to select all or one of the two acquisition modes.
The SpectrogramDecimationFactor, LowPassCutoffFrequency, HighPassCutoffFrequency device attributes are non-NULL for all icListen LF hydrophones. The SpectrogramDecimationFactor is probably misnamed, there is no decimation happening, instead, it increases the nFFT, which increases the number of frequency/y-axis bins (increasing spectral resolution, while decreasing temporal resolution). LowPassCutoffFrequency, HighPassCutoffFrequency attributes are also somewhat misnamed, they aren't associated with actual filters, instead, when not null, they act to trim off the low and high ends of the spectral data, reducing the range of the frequency/y-axis in the plot and in the spectral data MAT files. |
Below is an example of the latest version of the spectrogram data product:
The PDF format contains multiple pages, with each page containing one spectrogram. We recommend this format when users would like to scroll through a large amount data looking for events such as whale calls. PDF spectrograms are not normally archived for fast retrieval, so they will be generated on the fly, which will take some time. PDF spectrograms do have the advantage of having higher resolution, approximately 300 dpi for landscape letter sized image, while the PNG spectrograms are 1200 by 900 pixels.
Spectrograms can be produced from WAV files (described above) or from FFT files (described below). Spectrograms generated from FFT files have a fixed and generally lower resolution, but have the advantages of not being affected by military diversion and have wider frequency range (WAV-sourced spectrograms maybe limited in frequency by their multi-point calibration). Here is an example of a spectrogram generated from an FFT file:
The spectrogram collation option allows users to group/collate spectral data into plots and data (MAT) files. For plots, daily and weekly spectrograms are available, while the data files can be daily or unlimited duration collations. Spectral data for daily plots and also data files are assembled into 1-minute box car averages (no overlap), accounting for the logarithmic scale of the data. Spectral data for the weekly plots are assembled into 5-minute box car averages (no overlap), accounting for the logarithmic scale of the data. These plots are useful for daily and monthly inspections of data and, as such, will appear on Data Preview. Here's an example where a passing ship can be seen in a daily plot:
Here is an example of a weekly plot:
Oceans 3.0 API filter: extension={png,pdf}
The FFT format is an ASCII text file with a single column of data. It is intended for expert users, while other users may defer to the spectrogram PNG/PDF plots, which may be made from FFT files on user option or when WAV files (audio data) is not available (FFT files are available when audio data is diverted by the military). FFT files are only offered on icListen HF and AF hydrophones and often only for devices with low-bandwidth connections, such as the hydrophone currently at Cambridge Bay (deviceID [instruments:23155]). The file consists of repeating sequences of 512 FFT spectral coefficients, spanning five minutes. The current sampling rate is 256 kHz, with 4 FFTs per second, or 1200 in one file, with a frequency bin spacing of 250 Hz. Using MATLAB, one can visualize it (i.e. make a spectrogram) quickly with the following commands:
data = dlmRead('myFFTfile.fft', ',');
specData = reshape(data, [512, length(data)/512]);
imagesc((1:size(specData,2))/4, (511:-1:0)*0.250, flipud(specData));
axis xy
xlabel('Time (seconds)');
ylabel('Frequency (kHz)');
cb = colorbar;
ylabel(cb,'(dB re 1 \muPa)'); |
Please note the above stub of MATLAB code is an example only, with hard-coded parameters.
Oceans 3.0 API filter: extension=fft
The OCT format is an ASCII text file with a single column of data, very similar to the FFT format described above. Both FFT and OCT files represent sound intensity in the frequency domain. It is intended for expert users, while other users may defer to the spectrogram PNG/PDF plots. OCT files are only offered on specific JASCO / GeoSpectrum hydrophones. The file consists of repeating sequences of 55 1/3 octave band sound pressure levels in dB, with a single point calibration. The centre frequency of each 1/3 octave band is shown in the table within the expander below. The files nominally span five minutes and values are reported once per second, so a five minute file should contain 16,500 entries. One can visualize the data in Matlab with code similar to the code presented above for FFT files, but with modifications for the 55 bands and their frequencies. If you are interested in viewing this data please contact us, we can help develop visualization and perhaps a data product.
|
Oceans 3.0 API filter: extension=oct
The MAT file format is based on the same data used to create the spectrograms. By default, it contains spectral data that is resampled to one-minute average ensembles (the data is converted from dB to linear, averaged, converted back to dB, the frequency bins may also be downsampled so that the maximum number of bins is less than 2400). The one-minute average ensembles are used as source data for spectral probability density plots. These files are pre-processed and stored for fast retrieval. If pre-processed MAT files do not exist in the archive, then they are created on the fly, which is much slower. On retrieval or on-the-fly generation, there is one small, ~150 kB, MAT file per wave or hyd source audio file. For ease of use, the multiple small MAT files are then concatenated, with the concatenated MAT files breaking on configuration change (exceedingly rare) or on a size limit of approximately 1 GB in memory. The concatenation process applies a weighted average of ensemble periods that overlap between the small MAT files, accounting for count and conversion to linear scale and back to dB.
To directly access the data plotted in the spectrogram images, users may choose the "Spectrogram resolution" option noted in the Spectral Data Downsampling option, or the full resolution spectral data, the parameters of which are set by the calibration and hydrophone sample rate. The spectrograms are downsampled to match in the available pixels on the image - downsampling in this way is preferable to allowing the image plotting/rendering stage to do as our downsampling converts dB data to pressure/linear units, downsampling and then converts back (same as done for the one-minute ensembles). These MAT files are not stored so they take some time to generate and are not available to concatenate (one MAT file per spectrogram / source audio file), but they do have the same format as the one-minute average MAT files.
Hydrophone spectral data MAT files contain two structures, the nominal complex data metadata structure Meta and the data: SpectData
SpectData: structure containing hydrophone spectral data in the following fields:
Oceans 3.0 API filter: extension=mat
To comment on this product, log in and click _Write a comment..._ below.