New technology for the detection of micronekton:
multivariate acoustics, sampling and data analysis strategies
Maurice-Lamontagne Institute, Fisheries and Oceans Canada,
850 route de la Mer, P.O. Box 1000, Mont-Joli, Québec G5H-3Z4
What is micronekton?
Micronekton includes a large group of organisms, ubiquitously distributed in the ocean, which are defined by their size and their swimming capacity. This "fuzzy" definition makes micronekton mean different things in the mind of scientists. Depending on the point of view, micronekton can include, or not, macro-zooplankton species, such as krill and amphipods crustaceans, or some or all fishes up to a given size that varies from one person to the other. A working definition of micronekton (but see Pearcy 1983 and reference therein) is all mobile organisms within a size range. The lower size limit is given by truly planktonic animals that can not fight any significant horizontal current (e.g. copepods), and the upper limit is given by the large nekton that can fight large current speeds. Taking 1 knot (~50 cm s-1) as the cut-off current speed and four body length per second for the swimming speed an organism can sustain over prolonged periods, the micronekton upper size becomes ~12.5 cm. Micronekton thus ranges from ~1-2 cm (macro-zooplankton) to ~12.5 cm. This comprises, among others, all fish larvae (ichthyoplankton), juvenile fish, forage fish, small pelagic fish (including mesopelagic fish of deep sound scattering layers (DSLs)), krill, pelagic shrimps, and small squids. As plankton and fish, micronekton is structured in complex 3-D spatial patterns at all scales in the ocean, from metres to hundreds of kilometres (e.g. Figure 1).
Sampling micronekton has always preoccupied scientists because a series of specially designed instruments, varying from plankton nets to pelagic trawls, is required to adequately sample this size group of highly-mobile animals. Standard plankton nets cannot be used because of their generally small mouth opening and their avoidance up to certain degrees by most micronekton. Commercial pelagic trawls or other regular fishing gears cannot be used either, because of their too-large mesh size, which allows the escapement of most micronekton. The limited capacity to sample micronekton hindered the acquisition of knowledge on this group of organisms that plays an important role in the ecosystem, by transferring the energy from the lower trophic levels to the higher ones, which include all commercial fish and marine mammals. Efforts to find solutions to this problem were reviewed in the 1980s by an international group of specialists and reported in the proceedings of a symposium (Pearcy 1983). The advanced sampling tools then included single and multi-frequency acoustics combined with instrumented opening/closing nets and trawls, as well as cameras.
During the last two decades the advance in technology and knowledge considerably improved these tools, and some have now reached a reasonably good operational stage. Present technology to detect and estimate micronekton include a variety of acoustical and optical sensors, mounted on various platforms, combined with real-time monitored and controlled active and passive gears for directed physical sampling in the sound scattering layers or shoals. Among the optical tools are: low light level standard and video cameras, high resolution digital cameras, small cameras (that can be attached to a predator), multi-spectral cameras, LIDAR, bioluminescence sensors, activated bioluminescence coupled with cameras, and the optical particle counter (OPC).
The acoustic gears include high-resolution digital echosounders, single frequency to multi-frequency or wideband (range of ~30-200 kHz), single beam to multiple beams (dual-beam, split-beam, 2-D or 3-D multibeam sonars). The present trend is to use more frequencies and beams. The possibilities of low frequencies (<15 kHz) to detect micronekton and nekton over large ranges (> 1 km) have been explored. The acoustic gears also include acoustic tags that can be put on micronekton-feeding predators (e.g. tunas, whales). Fisheries acoustics made significant progress since 1980. The sound scattering properties of various taxa are much better known, as a result of better scattering models and target strengths (TS) measurements. Various data analysis methods and algorithms for classifying the echoes from their multivariate properties (echo statistics, echogram image analysis, spectral signatures combined with inversion methods, discriminant analysis, neural networks) have been developed and used for classifying single targets, schools and scattering layers. The statistical estimation methods also been improved to combine data to obtain objective estimates and 2-D or 3-D maps, notably by the use of the geostatistics and multivariate methods.
Physical sampling gears are now commonly equipped with various sensors to monitor the habitat characteristics (e.g. temperature, salinity, in situ light level, light transmission, fluorescence) and the instrument position and operation (depth, relative angle to the vessel, mouth opening, speed in water and filtering rate and efficiency, pitch and roll, catch detectors). The opening and closing of the net(s) is controlled from the surface in real-time or is time programmed. The sensors are generally transmitting data to the surface in real time, either via acoustical telecommunications (e.g. trawl monitoring systems) or via a cable (e.g. Bioness, Mocness).
Optical and acoustical micronekton detecting instruments have been installed on a variety of platforms besides the research vessels. These include: towed-bodies, buoys, moorings, fixed fishing gears, ROVs, AUVs, fishing vessels and submarines.
There are many challenges to efficiently make use of these new technologies in a 5 to 10 year period from now. Since the trend is to collect more information (for acoustics, see Horne, 1998), the co-ordination of the high-rate multi-channel data acquisition and output of data all streams, as well as the data management are serious challenges. The sampling and survey strategy is a non-negligible challenge. What to collect and how, requires thorough thinking. Converting the diverse and multivariate data output into meaningful integrated data streams represent a big data analysis challenge, especially when considering this has to be done for data sets collected over large areas and the results have to be properly validated. Other challenges are the data standardisation to facilitate their archiving and the exchanges of data and processing algorithms within the scientific community, and the enhancement of collaboration for the development of effective exportable solutions.
1. Sampling and survey strategies:
An effective survey design could be obtained from a simple adaptation of the traditional oceanographic approach consisting of sampling a regular grid of stations covering a study area where cross-sections (i.e. transects) of the water masses properties are sought. An operational strategy to detect, estimate and objectively map micronekton over large areas and depth ranges, as well as to simultaneously gather high-resolution data and physical samples to help identifying the sources of the echoes is shown in Figure 2. It uses the same high-resolution multi-channel acoustics sampling at the stations (high horizontal resolution) and on the track between the stations (continuous data over the whole water column). One channel can be the echo intensity at one acoustic frequency, for one beam, ADCP current velocities, or even processed data combining many channels of information. Physical samples of the scattering layers (SLs) can be collected at the oceanographic stations, as well as profiles of the environmental characteristics, sometimes in a single operation (e.g. from an instrumented net). High-resolution acoustic data is gathered at the same time with the same multi-channel instrument used on the track, in order to get more information to identify the echo sources and their TS (e.g. using the split beam technique on the numerous single target echo traces obtained as a result of the low the ship speed (e.g. Figure 9)). Additional (optical, acoustical, etc.) instruments can be to-yoed from the ship while sailing to the stations, to get higher resolution environmental and biological data along transects (e.g. Figure 5). The survey area can be flown over by an aircraft, possibly instrumented with optical instruments (e.g., multi-spectral camera, LIDAR), to localise aggregations of micronekton predators (e.g. tunas, dolphins, seals, and whales) and surface micronekton shoals.
Because of the limited penetration ranges of small acoustic wavelengths, the range sampled by general multi-frequency acoustics instruments aiming at micronekton (frequencies up to ~200 kHz) are limited to about 300 m. Therefore, to sample large depths, the instruments must be placed in an underwater vehicle (e.g. towed bodies, AUVs). Ideally, in addition to standard scientific capacity, the instrument characteristics must include for all frequencies, the attributes of sampling the same volume, having low detection thresholds (i.e. high signal to noise ratios (SNRs)), and producing high-quality data for echo amplitude, phase and spatial resolution in a standard and complete format. An example of a first step in that direction is the Fisheries and Oceans Canada CH* suite of tools (Simard et al. 1997, 1998, 2000) for multi-channel acoustic data collection and analysis (Figure 3). Such a tool proposes structured partial solutions to the above-mentioned challenge of co-ordinating the multi-channel dense data acquisition and management. The efforts to standardise the data format within the fisheries acoustic international community (Anon. 1999, Simard et al. 1997, 1999) is also a big step in the data management strategy. It should help to constitute high-quality large national and international fisheries and plankton acoustics data banks, which are needed for exchanges within the community and the development robust exportable solutions.
A survey strategy must include special designs to adequately sample the "hot spots" where intense micronekton-based activities occur, such as upwelling centres, frontal boundaries, and special topographic features. Satellite imagery helps to localise these hot spots, as remote detection of predators (e.g. tunas, marine mammals, fishing or whale watching fleet) from an aircraft or other detecting devices (e.g. detecting cetaceans from passive acoustics systems). However, different predators will point at different micronekton targets. An example of using such strategies for studying the semidiurnal dynamics at an intense whale watching spot in the St. Lawrence estuary (Marchand et al. 1999) is presented in Figures 4 and 5. A sampling track made of 5 transects (3 cross-shore and 2 long-shore) was surveyed repeatedly over the tidal cycle. The instrumentation included two-frequency acoustics to detect and sort out micronekton, continuous measurement of surface water physical properties, and measurement of optical (OPC) and physical properties on a to-yoed net while returning to the starting point via the central transect.
Survey design can also include moored sensors in strategic locations (e.g. shelf slope currents), to get long-term time-series of high-resolution multi-channel acoustics data under low noise conditions, in situ target and TS tracks with simultaneous images taken from a target-triggered camera and ADCP current data.
Ground truthing the acoustic information with micronekton physical samples or images can be done from various nets or optical samplers (see Pearcy 1983). However, all of them should be able to give the precise localisation of the sample relative to the acoustic data as well as the effective volume of water sampled. Ideally, to collect quantitative samples representative of the insonified volume, the samplers should not attract (e.g. with lights) the micronekton nor be avoided by some organisms. Also, the sample size (i.e. sample support in geostatistics) should be similar to the acoustic sample to which they are compared, a common problem in matching data from different sources. Ground truthing with physical samples in SLs require therefore real-time monitored opening/closing (either for mouth or codend) instrumented nets. Large multiple plankton nets (e.g. Bioness, Mocness) have these essential requirements to sample (with or without strobe lights) the lower part of the micronekton size range with a reasonable efficiency. To get physical samples for larger micronekton, small-mesh trawls (with uniform mesh?) must be used. Such gears generally include monitoring sensors connected to the surface with an acoustic link. They are, however, rarely equipped with flowmeters and remotely controlled opening/closing devices, even though they are required to properly associate the (uncontaminated) samples to the corresponding quantitative acoustic data. Only a few of such remotely controlled devices have been developed, and they are still at an experimental stage. Optical sampling devices (e.g. cameras, bioluminescence, OPC) offer possible alternatives to physical samples when they can produce precise images of the micronekton content of a significant volume unit (e.g. > 1m3), if they have the above-mentioned required attributes of valid ground truthing instruments. The same comment applies to the use of 3-D high-resolution sonars to get micronekton images (e.g. Jaffe et al. 1999). An example of ground truthing a micronekton scattering layer with multiple net samples from a Bioness tow is presented in Figure 6 (see Simard and Lavoie 1999).
Collecting ground-truthing and validation information is not exclusive to the above technology and methods. It can include additional approaches such as: strategically targeting the specific scattering layers that individually separate from the general DSL during the twilight periods of diel vertical migrations; sampling the areas intensively exploited by various predators, as shown by their distribution, or by following tagged predators (various VHF, acoustic and memory tags, including critter cameras); the use of fishing maps to match the results or to direct the sampling.
2. Data analysis:
After the acoustic data have been edited to remove any recorded noise or sound that does not come from biological scatterers (ship or instrument noises, ship wakes and air bubbles, marine mammals echoes or calls, etc.), to define the right bottom depth, and to correct for the right medium properties (sound speed and absorption), a clean multi-channel echogram of raw data is available for data processing along survey transects (Figure 7). This multi-channel echogram is a series of 2-D XZ images of the same cross-section of the study area, one image representing the backscattering at one acoustic wavelength for one acoustic beam or the result of a given processing (e.g. split-beam single target recognition and in situ TS estimates). More XZ plans of information are added to these primary acoustic data, by joining the environmental information on transects (ADCP, temperature, salinity, light levels, chlorophyll, particle concentrations, etc.). Then, extracted secondary XZ variables add new information layers to the data set. These could be echo statistics (e.g. image analysis outputs of local echo roughness on the echogram, of the mean amplitude in a given neighbourhood, of a given filter), or relations between the various data channels (e.g. ratio between the echo amplitude at two frequencies, results of a backscattering model or a multivariate analysis using the information from many channels). At this step a multivariate echogram is built, where each pixel is defined by a multivariate. This multivariate is used to sort the echo pixels into homogeneous categories (species or taxa assemblages) in combination with the information from the ground truthing data set and from the backscattering models. This echo identification proceeds as usual by feeding classification algorithms (cluster analysis, discriminant analysis, neural nets, etc.) with the multivariate description and training data set (ground truthing samples) to extract the various taxa "signatures". In this process, care must be taken to insure that the multivariate description is not deteriorating with echo range (e.g. because of the different absorption at the various frequencies, different SNRs).
Processing the multivariate echogram for echo classification should be done by step (Figure 7). First, the fish shoals (or schools, Figure 8) should be extracted first and processed by school recognition algorithms. Such image analysis based school classifications (e.g. Weill et al. 1993) have been used with reasonable success, at least on a regional and seasonal basis, in various regions of the world. Here, the information from a multibeam sonar system can help to bring more spatial information to identify the fish school. Second, the single target echo traces (Figure 9) should be handled with dedicated algorithms (single to multiple frequencies or beams) (e.g. Figure 10), to identify their most probable sources. This information could also help to identify the schools, if the single-target fishes surrounding the schools can be associated to them. Third, the rest of the multivariate echoes (Figure 11) should go to various other classification algorithms. Some simple ones are commonly used (e.g. sorting out krill from other echoes (e.g. Figure 12 to 14, see Simard and Lavoie 1999)), but most of the algorithms needed to objectively classify these other echoes are still to be developed. The use of the spectral signature combined with scattering models is one of the possibilities. The information from the previously classified schools and single targets can also contribute to the classification (Figure 7).
Once the multivariate echogram has been classified, the next data analysis step is to validate the results. Are the echo class and biomass spatial organisations (e.g. Figure 15) reasonable? given our knowledge of the local oceanography (water masses structure, currents, etc.)? Are the predator/prey organisation and biomass levels consistent with a functioning ecosystem? Do the observed patterns match with fishing maps? These are questions that a validation step should answer to strengthen the confidence in the derived results.
The last step of the data analysis process is to produce unbiased estimates and 2-D (Figure 16) or 3-D (Figure 17) maps of the biomass per species, or taxa assemblages depending on the objectives of the research project, with their confidence limits. Uni- or multivariate geostatistical methods, taking advantage of the spatial autocorrelation structure of the high-resolution acoustic data, are used here (e.g. Simard and Lavoie 1999, Lavoie et al. 2000). The anisotropy of the biomass field along the 3 XYZ dimensions must be taken into account as well as the cross-correlation with ubiquitous co-variate (e.g. topography, interpolated or satellite-derived images of temperature fields) to get optimal estimates and maps.
Finally, since micronekton is at the interface between zooplankton and commercially exploited large nekton, the technology and methodology used in both fisheries and plankton acoustics to detect and estimate the biomass are, to a large extent, applicable to micronekton. The technological development in these research fields thus contribute to the improvement of our capacity to efficiently detect and estimate micronekton.
Anon. 1999. Report of the working group on fisheries acoustic sciences and technology. ICES CM 1999/B:2
Horne, J.K. 1998. Acoustic approaches to remote species identification. In: Parrish, J.K., Remote species identification (RSID) workshop report. Census of Marine Life. University of Washington. 35 p.
Jaffe, J.S., A. De Robertis and M.D. Ohman. 1999. Sonar estimate of daytime activity levels of Euphausia pacifica in Saanich Inlet. Can. J. Fish. Aquat. Sci. 56: 2000-2010.
Lavoie, D., Y. Simard, and F.J. Saucier. 2000. Aggregation an dispersion of krill at channel heads and shelf edges: the dynamics in the Saguenay - St. Lawrence Marine Park. Can. J. Fish. Aquat. Sci. 57: 1853-1869.
Pearcy, W.G. 1983. (ed.). Methods of Sampling Micronekton. Biol. Oceanogr. Vol. 2, no. 2-4. (Proceedings of SCOR Symposium on Methods of Sampling Micronekton, Working Group 52, held in April 1980 at Idyllwild Ca.)
Simard, Y. and D. Lavoie. 1999. The rich krill aggregation of the Saguenay - St. Lawrence Marine Park: hydroacoustic and geostatistical biomass estimates, structure, variability and significance for whales. Can. J. Fish. Aquat. Sci. 56: 1182-1197.
Simard, Y., I. McQuinn, N. Diner, and C. Marchalot. 1999. The world according to HAC: summary of this hydroacoustic standard data format and examples of its application under diverse configurations with various echosounders and data acquisition softwares. ICES-Fisheries Acoustics Sciences and Technology meeting, St. John's, Newfoundland, Canada, 20-22 April 1999, Working paper. 14 pp.
Simard, Y., I. McQuinn, M. Montminy, C. Lang, D. Miller, C. Stevens, D. Wiggins and C. Marchalot. 1997. Description of the HAC standard format for raw and edited hydroacoustic data, version 1.0. Can. Tech. Rep. Fish. Aquat. Sci. 2174: vii + 65 pp.
Simard, Y., I. McQuinn, M. Montminy, C. Lang, C. Stevens, F. Goulet, J.-P. Lapierre, J.-L. Beaulieu, J. Landry, Y. Samson and M. Gagné. 2000. CH2, Canadian hydroacoustic data analysis tool 2 user's manual (version 2.0). Can. Tech. Rep. Fish. Aquat. Sci. 2332: vii + 123 pp. (in press)
Simard, Y., I. McQuinn, M. Montminy, Y. Samson, C. Lang, C. Stevens and D. Miller. 1998. CH1, Canadian hydroacoustic data analysis tool 1, user's manual (version 2.0). Can. Tech. Rep. Fish. Aquat. Sci. 2256: viii+ 100 pp.
Weill, A., C. Scalabrin and N. Diner.1993. MOVIES-B: an acoustic detection description software. Application to shoal species classification. Aquat. Living Resour. 6: 255-267.
Figure 1. Example of marine life spatial structures in an ocean voxel from the Gulf of St-Lawrence, as depicted by 38 kHz sound backscattering. Most of the weakly backscattering structures (green and blue) at this acoustic wavelength are from micronekton, as define above. (The strongly sound backscattering structures (orange and red) are from an Atlantic cod migrating shoal).
Figure 2. Example of an adaptation of the systematic oceanographic survey design for the detection, estimation and objective mapping of micronekton using the multivariate acoustic approach.
Figure 3. Fisheries and Oceans Canada CH* suite for multi-channel acoustic data processing and archiving under the standard international HAC data format.
Figure 4. Sampling circuit repeatedly visited over a semidiurnal tidal cycle to intensively survey a micronekton "hot spot" in the St. Lawrence estuary with multi-frequency acoustics. Over a complete circuit, 3 cross-channel transects are surveyed as well as 2 transects along the channel borders. The central transect is surveyed twice. On the second pass, it is simultaneously sampled with a to-yoed instrumented net.
Figure 5. Examples of the data resulting from the central cross-channel transect of the sampling circuit shown on Figure 4.
Figure 6. Examples of ground truthing micronekton sound scattering layers with physical samples from an instrumented net. The paths sampled by the Bioness nets are somewhere in the in white rectangles on the binned echogram. The catches for each net and depth interval are presented as bar graphs in the right panel.
Figure 7. Multivariate acoustics data analysis process for 2-D and 3-D estimation and mapping of micronekton species and taxa assemblages according to research objectives.
Figure 8. 38 kHz echogram from the St. Lawrence estuary, showing many micronekton target tracks dispersed over a large depth layer between 20 and 90 m. A detailed analysis of one of these tracks is presented in Figure 10.
Figure 9. Example of processing single targets detected with the split-beam technique from the 38 kHz echogram shown in Figure 9. The individual tracks are automatically identified according to contiguity attributes of the single target echo time-series, the consistency in the beam-corrected TS, and some criteria about the 3-D target trajectory and swimming speed. Here the red-dot track crossing the vertical red line on the left is positioned relative to the transducer and acoustic beam in the vertical and horizontal planes in the front window (upper graphs) as well as in time (lower left graph). The (consistent) beam-corrected TS for each measurement of the target is presented in the red bar graph. The statistics for the track are presented in the lower right panel. The probability density functions of these statistics for all tracks are then computed to get the average TS, and the angle aspect distribution of the tracks in the vertical and horizontal planes.
Figure 10. Complement image of the 38 kHz echogram shown in Figure 8 (i.e. Figure 13 minus Figure 8), showing the "other micronekton" when the series of small micronekton shoals is removed. The classification of this "other micronekton" echogram must follow a different algorithm than the shoal classification algorithm. An example of the classification of the large scattering layer below 50 m with a two-frequency algorithm is presented in Figure 14.
Figure 11. Scatter plot showing the difference in volume backscattering strengths at 120 kHz and 38 kHz for ground truthed samples of krill and micronekton fish from the St.Lawrence estuary (see Simard and Lavoie 1999).
Figure 12. The same 38 kHz echogram of Figure 8 but with the echo signal truncated at a lower threshold level, similarly to those used in zooplankton acoustics, showing the whole micronekton picture (essentially capelin and krill).
Figure 13. Example of the use of the two-frequency signature shown in Figure 12 to classify the krill in the scattering layer topped by fish schools (shown in Figures 11 and 13) below the 20th layer of vertical bins. Binned echograms in small voxels as indicated. The axes indicate the bin co-ordinates (not depth and ping sequence). Note the clear krill signal in red in the scattering layer in the differential echogram (bottom image) and the fish and krill mix (blue-green) in the shoaling area at left, that appeared as a regular krill scattering layer on a single-frequency echogram (see Figure 11 or 13).
Figure 14. Micronekton acoustic data analysis validation step. Is the 3-D spatial organisation of the echo classes consistent with our understanding of the water mass structure, hydrodynamic processes and predator/prey interactions? Here, the picture obtained for the St. Lawrence estuary is consistent with the water mass structure, the krill being located in its preferred the cold intermediate layer, and the predator/prey hydrodynamically-controlled interaction model, in which the micronekton predator (capelin) is concentrated over the shallow areas at the channel head where the intense tidal upwelling is occurring and advects preys (see Lavoie et al. 2000).
Figure 15. Example of estimating and 2-D mapping two micronekton classes with linear geostatistics in the St. Lawrence estuary. The survey design and the 1319 km2 estimation area are shown in Figure 2. The measured 120 kHz area backscattering coefficient (m2/m2) on transects is transformed to kg per m2 from published relation of TS vs length and weight vs length. These data then serve to krige the biomass at each node of a 1 km2-mesh grid covering the study area, taking into account the cross-shore/along-shore anisotropy. The map is integrated to get the global biomass estimate, whose confidence interval is obtained from the spatial autocorrelation model and the geostatistical theory (see Simard and Lavoie 1999).
Figure 16. Example of estimating and 3-D mapping a krill aggregation over a segment of the St. Lawrence estuary. Same data and methods as in Figure 16, except that the volume backscattering coefficient (m2/m3) is used instead of the area backscattering coefficient. The spatial autocorrelation model is different. It is three-dimensional and therefore more complex. The local estimates are produced at each node of the 1319 km2 grid and each depth interval of 5 m, which represent a total of 41736 voxels of 1 km2 x 5 m filling the basin. The points representing each voxel is then coloured according its relative biomass per m3, as indicated. The vertical axis is the depth in metre and the horizontal axes are the Lambert projection of latitude and longitude co-ordinates, in km.
Figure 17. Example of estimating and 3-D mapping a krill aggregation over a segment of the St. Lawrence estuary. Same data and methods as in Figure 16, except that the volume backscattering coefficient (m2/m3) is used instead of the area backscattering coefficient. The spatial autocorrelation model is different. It is three-dimensional and therefore more complex. The local estimates are produced at each node of the 1319 km2 grid and each depth interval of 5 m, which represent a total of 41736 voxels of 1 km2 x 5 m filling the basin. The points representing each voxel is then coloured according its relative biomass per m3, as indicated. The vertical axis is the depth in metre and the horizontal axes are the Lambert projection of latitude and longitude co-ordinates, in km.
This page is maintained and was last updated 21-Nov-00 by Pip Sumsion (email@example.com).