Appl. Statist. (2011) 60, Part 3, pp. Estimating the density of Antarctic krill (Euphausia superba) from multi-beam echo-sounder observations using distance sampling methods Martin J. Cox and David L. Borchers, University of St Andrews, UK David A. Demer and George R. Cutter Southwest Fisheries Science Center, La Jolla, USA and Andrew S. Brierley University of St Andrews, UK [Received November 2009. Revised July 2010] Summary. Antarctic krill is a key species in the Antarctic food web, an important prey item for marine predators and a commercial fishery resource. Although single-beam echo-sounders are commonly used to survey the species, multi-beam echo-sounders may be more efficient because they sample a larger volume of water. However, multi-beam echo-sounders may miss animals because they involve lower energy densities. We adapt distance sampling theory to deal with this and to estimate krill density and biomass from a multi-beam echo-sounder survey. The method provides a general means for estimating density and biomass from multi-beam echo-sounder data. Keywords: Acoustic; Antarctic krill; Distance sampling; Multi-beam echo-sounder 1. Introduction Antarctic krill (Euphausia superba; hereafter simply called krill) are a key component in the Antarctic food web (Atkinson et al., 2001). The abundance and swarming behaviour of krill make it important to many marine predator species and as a commercial fishery resource; swarms can occur with thousands of individuals per cubic metre of sea and contain thousands of kilograms of krill. Management of the competing interests of marine predators and fisheries requires regular monitoring of krill abundance. Ship-based remote sensing acoustic surveys, which are conducted by using single-beam echo-sounders (SBEs), are key to this, with surveys frequently covering large areas (Brierley et al., 1999). Echo-sounders function by transmitting a sound pulse with known characteristics via a transducer into the column of water and recording the characteristics of the sound, or echo, that is scattered back towards the transducer. The intensity of the backscattered acoustic echo can be scaled by the use of appropriate acoustic target strength models (e.g. Demer and Conti (2005) for krill) to estimate the density of animals in the volume of water that is sampled by the echosounder. Modern scientific SBEs operating at acoustic frequencies of 120 and 200 kHz typically Address for correspondence: Martin J. Cox, Pelagic Ecology Research Group, Gatty Marine Laboratory, University of St Andrews, St Andrews, KY16 8LB, UK. E-mail: [email protected] © 2011 Royal Statistical Society 0035–9254/11/60000 2 M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley have a narrow beam width of 7◦ . In contrast, multi-beam echo-sounders (MBEs) have much wider swath widths (in this case 120◦ ) that are often made up of more than 100 narrow beams (individual beam widths in this case of 1:5◦ × 2◦ ) and so sample a much larger volume of water. This is particularly useful when studying organisms with aggregative behaviour, such as krill, since entire swarms can be acoustically sampled (see Gerlotto et al. (1999)). Methods that are used to estimate krill density from SBE data typically assume that all swarms within the SBE beam are detected (e.g. Brierley et al. (1999) and Reiss et al. (2008), but see Demer et al. (1999)). This may be a reasonable assumption in the case of SBEs because of the concentration of acoustic energy in a narrow beam. It is a less tenable assumption in the case of MBEs, where the acoustic energy density in the wider swath width is substantially lower. Swarms at greater distances from the transducer may be missed and, in this case, a model for the probability of detecting a swarm as a function of distance is required for unbiased density estimation. In this paper we extend and apply distance sampling methods to estimate the probability of detection and density of krill from MBE data. An extension of standard distance sampling theory is required because the assumption of standard theory that targets (swarms) are uniformly distributed fails in the acoustic survey context. In Section 2, we describe the key data obtained from the MBE survey. We summarize standard distance sampling theory and consider the validity of its key assumptions in the context of acoustic surveys in Section 3. Section 4 describes extensions to the theory to deal with the failure of a key assumption of standard theory on acoustic surveys and develops the associated likelihood function and maximum likelihood estimator. The bias of the maximum likelihood estimator is investigated by simulation in Section 5. The method is applied to estimate krill density and biomass by using data from an MBE survey off the South Shetland Islands in Section 6. 2. Data Krill swarms were sampled by using an SM20 MBE (acoustic frequency 200 kHz; Kongsberg Mesotech Ltd, Vancouver, Canada). MBE data were collected along 41 line transects, lengths varying from 2.5 to 3.5 km, from February 2nd–9th, 2006, near Cape Shirreff, Livingston Island, South Shetland Islands, Antarctica (Fig. 1). Krill swarm boundaries were identified from the data by using the acoustic processing software (Echoview v3.5, Myriax, Hobart, Australia). Key data that are used in the analysis are shown in Table 1 and Fig. 2. The MBE mean volume backscattering strength Sv (see MacLennan et al. (2002) for a formal definition) observations were calibrated by using the technique that is described by Cox et al. (2010) and scaled by using the estimate of krill target strength from the krill sound scattering model of Demer and Conti (2005), from which swarm biomass b was estimated. (Sv is the sum of the acoustic backscattering area (square metres) of organisms, in this case krill, per unit volume (cubic metres), in the logarithm domain, dB re 1 m−1 .) In common with many acoustic survey analyses (e.g. Brierley et al. (1999) and Reiss et al. (2008)), we treat individual swarm biomass as being estimated without error. A total of n = 1006 krill swarms were detected on the survey. 3. Standard distance sampling Distance sampling surveys are used to estimate abundance N or density D of objects (which are called ‘animals’ henceforth) when some animals in the region searched are missed and the probability of detection depends on the distance from the observer. The key assumptions Estimating Density of Antarctic Krill 3 Fig. 1. Cape Shirreff study site: line transects are given in grey scale; C, detected krill swarm geometric centres Table 1. Description of swarm location data used in this analysis† Measurement Swarm centre horizontal distance Swarm centre depth Swarm centre radial distance Swarm centre angle Truncation distance Swarm biomass Symbol Units x y r θ w b m m m rad m kg †Swarm centre location is measured relative to the MBE. of standard distance sampling methods are as follows (from Buckland et al. (2001), pages 18–19). (a) Survey lines or points are randomly placed with respect to animals, or systematically placed with a random start point. (b) All animals at distance 0 are detected. (c) Animals are detected at their initial location. (d) Measurements are exact. The two main types of distance sampling survey are line transect and point transect. In the case of line transect surveys, observers move along each of a set of lines and survey a twodimensional region within some distance w of the lines; in the case of point transect surveys, observers survey a fraction of a two-dimensional circle about each of a set of points. Although 4 M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley Fig. 2. MBE swath geometry and location measurements .ri , θi / for the i th krill swarm detected (one half of the swath is shown; larger acoustic volume backscattering strength values are shown as darker samples; data at radial distances beyond w are truncated; y Å is the depth at which the width of the swarm is a maximum in the x -dimension; y D W sin.π=3/; , rectangle for which the expected probability of detecting a krill swarm is estimated (see equation (3), Section 4)): 1, swath vertical axis; 2, krill swarm with the boundary of the swarm delineated by using acoustic processing software (, geometric centre of the swarm); 3, seabed; 4, sea surface (x -axis); 5, maximum swath observation angle .θ D π=3/ echo-sounder surveys involve an ‘observer’ (the echo-sounder) traversing lines, for distance sampling analysis these surveys are more conveniently thought of as point transect surveys sampling in depth and perpendicular-to-transect dimensions (compressed along the length of the transect). The MBE samples a .2π=3/-rad segment of a circle about a point spanning the length of the line (see Fig. 2). Although assumptions (b)–(d) are defensible for MBE surveys, krill are known to exhibit a non-uniform vertical distribution (Demer and Hewitt, 1995; Demer, 2004) and so assumption (a) is probably flawed. This is true even when vessel transects are located randomly on the surface—because the transects are not located randomly with respect to depth. Assumption (a) enables the distribution of animals in two-dimensional space to be treated as known and uniform. Without this assumption the probability of detection g.r/ and distribution of animals with respect to distance from the observer πr .r/ are confounded. To see this, consider the point transect likelihood function for detected animals. Given radial distances r1 , . . . , rn to the n detected animals, the likelihood for φ, the unknown parameter vector of g.r/, is (from Buckland et al. (2001)) L.φ/ = n i=1 g.ri / πr .ri / w 0 , .1/ g.r/ πr .r/ dr where w is the maximum search distance (which in our case is the seabed depth—see Fig. 2). Because g.r/ and πr .r/ occur only as a product in the likelihood, they are not estimable separately. Hence standard distance sampling methods require πr .r/ to be known to estimate φ and Estimating Density of Antarctic Krill 5 g.r/. In the next section, we describe an extension of standard methods which allows separate estimation of πr .r/ and g.r/. Given πr .r/ and an estimator ĝ.r/, the standard point transect estimator of animal density from a survey in which a fraction γ of the circle is searched is D̂2 = n=as P̂ a , where w P̂ a = γ ĝ.r/ πr .r/ dr 0 is the expected probability of detecting an animal given that it is within a distance w of the observer (see below), and as = 2πw2 . In our context, D̂2 is an estimate of total density throughout the column of water in the along-surface (x) and depth (y) dimensions. 4. Distance sampling for multi-beam echo-sounder surveys As noted above, krill do not distribute themselves uniformly with respect to depth. To develop an estimator which allows for this and the possibility that the probability of detection is a function of angle θ, we write the detection function as p.r, θ/ and consider the joint probability density function (PDF) of radial distance r and angle θ of krill swarms, which can be written as πr,θ .r, θ/ = πr .r/ πθ|r .θ|r/. Note that the πs here are function names, not the mathematical constant. In common with standard distance sampling methods, we ‘fold’ the data about the vertical (θ = 0) for detection function estimation. In this case, the expected probability of detecting an animal within distance w of the MBE which searches out to an angle of γπ rad either side of the vertical can be written as γπ w Pa = p.r, θ/ πr,θ .r, θ/ dr dθ: .2/ 0 0 In the conventional distance sampling scenario p.r, θ/ =g.r/ and πθ|r .θ|r/ = 1=π (rememberw ing that we have folded about the vertical), so that Pa = γ 0 g.r/ πr .r/ dr, as given above, with 1 γ = 3 (see Fig. 2). In our case, because we want to model a non-uniform krill depth (y) distribution, it is convenient to work in terms of .x, y/ rather than .r, θ/. By locating transects randomly on the sea surface, we can treat the PDF of the horizontal distances from the line, given that they are within w sin.π=3/ of the line, as known and equal to 1=w sin.π=3/ (see Fewster et al. (2008) and Buckland et al. (2001), pages 232–235). We denote this PDF πx .x/. Random transect placement also means that we can treat y as independent of x and write πx,y .x, y/ = πy .y/=w sin.π=3/. Here πy .y/ is some smooth PDF with unknown parameter vector ϕ (see details in Section 4.4). It is now convenient to consider the estimation of density within the rectangle that is defined by 0 x w sin.π=3/ and 0 y w (see the broken line in Fig. 2) rather than the ‘pie slice’ 0 r w and 0 θ γπ. To deal with the fact that no detections are made at radial distances r > w and angles θ > γπ, we define p.r, θ/ to be 0 for r > w and for θ > γπ. We can then write the expected probability of detecting a krill swarm in the rectangular area as P= 0 w w sin.π=3/ 0 √ p{ .x2 + y2 /, tan−1 .x=y/} πy .y/ 1 dx dy: w sin.π=3/ .3/ Given models for p.r, θ/ with unknown parameter vector φ and for πy .y/ with unknown parameter vector ϕ, and that n krill swarms with centres .x1 , y1 /. . . , .xn , yn / were observed, the likelihood for φ and ϕ is 6 M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley L.φ, ϕ/ = √ n p{ .x2 + y2 /, tan−1 .x =y /} π .y /={w sin.π=3/} i i y i i i : P i=1 .4/ Although the above formulation is slightly different from that of Marques et al. (2010), the central ideas are taken from that work. 4.1. Random seabed depth The seabed depth varied along the transect line. Observations of seabed depth were available every 100 m, generating 1065 depth observations within the survey region. We used these observations to model the probability that the sea at a randomly chosen location was no less than y deep. We refer to this logistic function a.y, w/ as the depth attenuation function, and we use this function to account for the non-uniform seabed (Fig. 1) throughout the survey region. It models the proportion of the total transect length on which the depth was at least y. We model it as [1 + exp{−.y − β0 /=β1 }]−1 a.y, w/ = .5/ T.w/ where T.w/ = w [1 + exp{−.y − β0 /=β1 }]−1 dw, 0 and β0 and β1 are parameters to be estimated from the seabed depth observations. This functional form for a.y, w/ was chosen because it is bounded between 0 and 1, as is required for a cumulative distribution function, and it proved to be adequate. We found that models for πy .y/ with w equal to the maximum seabed depth provide poor fits to the data. Because w in the model corresponds to a depth below which no krill can occur (rather than bottom depth per se), we also considered models in which we estimated w. In these models, we constrained w to be between the maximum observed krill swarm depth and maximum seabed depth. Inclusion of a.y, w/ leads to the following likelihood in place of equation (4): √ n p{ .x2 + y2 /, tan−1 .x =y /} π .y /.1=x i i y i max / a.yi , w/ i i L.φ, ϕ, w/ = .6/ Å P i=1 where PÅ = 0 w xmax 0 √ p{ .x2 + y2 /, tan−1 .x=y/} πy .y/.1=xmax / a.y, w/ dx dy, and xmax is the maximum rectangle width in Fig. 2, which is equal to w sin.π=3/. We maximize the likelihood equation (6) to obtain maximum likelihood estimates of φ, ϕ and w (taking β0 and β1 as known). 4.2. Krill density, abundance and biomass estimators We estimate the number of krill swarms per unit volume of sea as D̂s3 = n 1 L 2xmax ŵP̂ Å Å where P̂ is P Å evaluated at the maximum likelihood estimators φ̂, ϕ̂ and ŵ. .7/ Estimating Density of Antarctic Krill 7 Krill swarm areal density (the number of swarms per unit area of sea surface) is estimated by D̂s = D̂s3 ŵ, and the krill swarm abundance in the survey region (with surface area A) is estimated by N̂ s = AD̂s . In standard distance sampling, estimates of individual density and abundance are usually obtained by multiplying estimates of group density or abundance by estimated mean size of group. Because there may be size selectivity in observing groups, the mean size of group is conventionally estimated by regressing log(group size) against distance (or ĝ.x/), and using the value of the regression line at distance 0 (where all groups are seen, by assumption). See Buckland et al. (2001), pages 73–75, for details of the method. We use a similar method to estimate the mean biomass of the swarm and total biomass in the survey region. We expect the mean biomass of observed swarms (the mean of b) to decrease at larger angles θ because the probability of detection decreases with θ (see Section 6) and because parts of swarms at larger θ may be undetectable. We therefore regressed the logarithm of observed swarm biomass log.b/ against θ and used the estimated intercept α̂ and residual variance σ̂ 2 from the regression to estimate the mean biomass of the swarm and to correct for bias (Ê[b] = exp.α̂ + σ̂ 2 =2/, where b is the biomass of the swarm). A regression of log.b/ against radial distance r was found to have zero slope. The total krill biomass in the survey region was estimated by B̂ = N̂ s Ê[b] and the biomass density was estimated by ρ̂ = B̂=A. Variance and 95% confidence intervals were estimated by the non-parametric bootstrap with 1000 replicates, using a transect as the sampling unit. Confidence intervals were calculated by using the percentile method. The survey contained 41 transects of varying length and the transect selection probability in the bootstrap was made proportional to transect length. To use the above estimators, we need suitable models for the detection function p.r, θ/, and the krill depth distribution PDF πy .y/. We develop these in the next two subsections. 4.3. Krill detection function models The probability of detecting a swarm is modelled as p.r, θ/ = g.r/ q.θ/. Standard distance sampling detection functions can be used for g.r/; we use the half-normal form, with a single parameter φ1 : 2 −r g.r/ = exp : .8/ 2φ21 Without independent data on q.θ/, it and πy .y/ are confounded. Generally, MBE systems are less likely to detect an acoustic target in the outer beams (those at larger θ) than in the centre beams. The beam-by-beam sensitivity was estimated by using observations of MBE beam-bybeam noise. Noise was assumed to be isotropic and estimated by using Sv -data that were collected at the Cape Shirreff study site. The beam-by-beam logarithmic Sv (see for a definition MacLennan et al. (2002)) observations with the MBE operating in ‘passive mode’ and transformed to the linear domain (Sv = 10Sv =10 ) and inverted so that beams with lower noise are more likely to detect krill swarms, giving a relative measure of krill swarm detection probability. The passive mode Sv -data were collected within the survey region, but not during the krill survey. The following hazard rate functional form with parameters α0 , αθ and α1 was found to be adequate: −α1 θ : .9/ q.θ/ = α0 1 − exp − αθ 8 M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley 4.4. Krill depth distribution model The density of krill swarms is known to be very low at the surface (Brierley et al., 1999) and is expected to be low beyond 100 m (Demer, 2004): (a) a left- and right-truncated normal distribution, y|w ∼ N.ϕ1 , ϕ22 /={F.w/ − F.0/}, where F.w/ is the cumulative distribution function of y evaluated at w and F.0/ is the cumulative distribution function evaluated at 0; (b) a right-truncated beta distribution, πy ϕ2 −1 y y Γ.ϕ1 + ϕ2 / y ϕ1 −1 1− ; w = w Γ.ϕ1 / Γ.ϕ2 / w w (c) a right-truncated log-normal distribution, log.y/|w ∼ Nw .ϕ1 , ϕ22 /={F.w/ − F.0/}, where F.w/ is the cumulative distribution function of log.y/ evaluated at w and F.0/ is the cumulative distribution function evaluated at 0; (d) a uniform distribution: πy .y|w/ = 1=w. 5. Simulation testing We conducted simulations to investigate the bias of the maximum likelihood estimator of abundance, simulating from a uniform πx .x/, the normal model for πy .y/ and the half-normal model for g.r/ that were fitted to the survey data (see Section 6). Simulations were conducted for a range of krill swarm abundances in the pie slice of Fig. 2 (Nsim ), ranging from Nsim = 100 to Nsim = 2200, in increments of 50. For each Nsim 1000 simulations were performed. The simulated sampling distribution of N̂ sim and its mean N̂¯ sim is plotted against Nsim in Fig. 3(a). The relative bias (.Nsim − N̂ sim /=Nsim , where N̂ sim is the estimate for a given Nsim ) is shown in Fig. 3(b). Fig. 3 indicates that the abundance estimator is approximately unbiased for Nsim greater than 500, which corresponds to an expected sample size of about n = 250. At smaller sample sizes it is slightly positively biased. Fig. 3(a) also suggests that the estimator has a substantially more skewed sampling distribution than a normal random variable, except perhaps for large Nsim (2200 or more), corresponding to a sample size n of more than about 1200. 6. Results 6.1. Seabed depth distribution model The distribution of observed depths on the survey is shown in Fig. 4(a). The complement of the empirical distribution function of depths .1 − CDF.y/, the observed proportion of depths greater than y), together with the estimated probability that a depth is greater than y .a.y, w/; see equation (5)) is shown in Fig. 4(b). The model provides a very good fit. 6.2. Angular detection function model Data collected during the SM20 calibration exercise that was conducted at Cape Shirreff showed an increased background noise on beams towards the edge of the MBE swath, as shown in Fig. 5(a). The model for q.θ/ that is given in equation (9) provided a good fit to the inverse of background noise (Fig. 5(b)) and showed very little loss in detectability at angles that are less than θ = 0:7 rad, after which detectability dropped off rapidly. 9 1500 1000 0 500 ^ Estimated abundance (Nsim) 2000 2500 Estimating Density of Antarctic Krill 0 500 1000 1500 2000 2500 0.4 Simulated abundance (Nsim) (a) ● ● ● ● 0.0 0.2 ● ● –0.2 ^ Relative bias (Nsim−Nsim)/Nsim ● ● –0.4 ● 200 400 600 800 1000 1250 1500 1750 2000 Nsim (b) Fig. 3. Simulated sampling distribution of N̂ s over a range of known swarm abundances (Nsim D 200–2200 in increments of 50; there were 1000 simulations for each Nsim ): (a) features of the simulated sampling distribution of N̂ sim ( ) plotted against Nsim , with the mean N̂¯ sim ./ and ˙2 standard errors (5, lower bound; 4, upper bound) for each p Nsim -value; (b) boxplots of the relative bias, .N̂ sim Nsim /=Nsim (the pair of curves are at 1:96 s e.N̂ sim /= 1000 above and below the mean N̂ sim ) 1.0 M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley 0.6 0.4 200 0 0.0 0.2 100 Frequency 300 0.8 400 10 40 60 80 y, (a) 100 120 0 50 100 150 (b) Fig. 4. Estimated distribution of seabed depths, a.y/: (a) distribution of mean seabed depths over each 100-m segment of the transect line; (b) proportion of depths greater than y (- - - - - - - ) and the estimated a.y/ (see equation (5)) fitted to the data, β̂ 0 D 0:115 (coefficient of variation 18%) and β̂ 1 D 5:920 (coefficient of variation 0.03%) 6.3. Vertical distribution and detection function models Parameters of the krill vertical distribution PDF πy .y/ and radial distance detection function g.r/ were estimated by maximizing likelihood equation (6) with respect to φ, ϕ and w, treating â.y, w/ and q̂.θ/ as known. The krill vertical distribution model was selected by using Akaike’s information criterion AIC and the goodness of fit was assessed by using a χ2 -test with the depth intervals that are shown in Fig. 6. Expected values in each depth interval were obtained by integrating under the full curve in Fig. 6. The truncated normal model for πy .y/ was selected on the basis of AIC-weight (0.984) and was the only model with an acceptable goodness of fit. See Table 2 and Fig. 6. For all πy .y/, the likelihood that maximized φ, ϕ and w consistently outperformed the likelihood that maximized only φ and ϕ (equation (4)), which had both higher AIC-values and goodness-of-fit, χ2 , that were significant for all πy .y/. 6.4. Estimated krill density and biomass Density, abundance and biomass estimates are shown in Table 3, together with the encounter rate and πy .y/ and g.r/ parameter estimates. The regression from which the mean swarm biomass (E[b]) was estimated is shown in Fig. 7. 7. Discussion Investigations of the spatial distribution of krill at the South Shetland Islands during the austral summer have shown that krill undertake diel vertical migration (DVM) (see Demer and Hewitt (1995) and Demer (2004)) and reside mostly in the upper 40 m, making it extremely unlikely that krill swarms have a uniform vertical PDF. At larger temporal scales to the west of the Antarctic Peninsula, seasonal variation in krill biomass has been shown, with krill biomass located in 11 26 28 θ 20 22 24 Sv m−1) Estimating Density of Antarctic Krill 0 20 40 60 80 100 120 1.2 0.2 0.4 0.6 0.8 1.0 θ) 1.4 (a) 0.0 0.2 0.4 0.6 θ 0.8 1.0 (b) Fig. 5. Variation in beam-by-beam sensitivity of the SM20 MBE that was used during the krill survey (the model is given in equation (9); α̂0 D 1:12 (coefficient of variation 1.6%); α̂0 D 0:87 (coefficient of variation 1%); α̂1 D 10:37 (coefficient of variation 14.1%)): (a) background noise, which increases (higher Sv ) towards the edges of the swath; (b) inverse of the data in (a) (), folded about θ D 0, and the fitted angular detection function q̂.θ/ ( ) M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley 0 0 50 50 100 100 150 150 200 200 250 250 300 300 12 0 20 40 60 80 20 40 80 100 60 80 100 0 0 50 50 100 100 150 150 200 200 250 250 60 (b) 300 (a) 300 0 100 0 20 40 60 (c) 80 100 0 20 40 (d) Fig. 6. Model fits for various krill vertical distribution models (- - - - - - - ) (πy .y/, with parameter estimates ϕ̂1 and ϕ̂2 ) and half-normal radial distance detection function model (. . . . . . . ) (g.r/, with parameter φ1 ) (Ŵ D 100 m for all models; , expected number of detections, conditional on the total sample size; πy .y/ and g.r/ have been scaled to correspond to the scale of the expected number of detections): the histograms show the observed frequencies of the swarm in 10-m depth bins and the vertical distribution models are (a) normal (ϕ̂1 D 61:1 m; ϕ̂2 D 27:8; ϕ1 D 76:6 m), (b) beta (ϕ̂1 D 2:27; ϕ̂2 D 1:70; ϕ1 D 78:6 m), (c) log-normal (ϕ̂1 D 5:23; ϕ̂2 D 0:99; ϕ1 D 71:4 m) and (d) uniform (ϕ̂1 D 99:8 m) (the normal distribution was selected on the basis of AIC deeper water in the winter, and shallower in the summer (Lascara et al., 1999). Non-uniform krill vertical distributions have been found at other locations also, e.g. around South Georgia (54◦ S 35◦ W), where krill have been shown to undertake DVM that varies with location relative to the continental shelf (Cresswell et al., 2007). It would be unrealistic to assume a uniform krill swarm vertical distribution in general, but given the high variation in krill vertical distribution it would also be unrealistic to prespecify a depth distribution model. Estimating the krill vertical distribution is particularly important for estimating the accessibility of krill to air breathing predators: many krill predators are constrained in their depths of diving, so being able to stratify the krill density estimates by depth will facilitate estimation of biologically plausible prey fields. Furthermore, estimating krill swarm density will enable Estimating Density of Antarctic Krill Table 2. 13 Parameter estimates† Krill vertical distribution πy (y) ϕ̂1 g(r), πy (y), ŵ φ̂1 AIC ΔAIC AIC-weight χ2 p-value ϕ̂2 Normal 61.1 27.8 76.6 Beta 2.27 1.70 78.9 Log-normal 5.23 0.99 71.1 Uniform — — 126.4 100 100 100 100 7899.4 7907.7 7935.8 7996.1 0 8.3 36.4 96.7 0.984 0.016 0 0 0.2 0.02 0 0 †ΔAIC is the difference in AIC from the model with smallest AIC. The χ2 p-value is for a χ2 goodness-of-fit test. Table 3. region† Model parameter and krill biomass variance estimates for the survey Parameter Normal ϕ̂1 Normal ϕ̂2 Half-normal φ̂1 ŵ Encounter rate n=L (km−1 ) Å P̂ N̂ s Ê[b] (kg) B̂ (tonnes) ρ̂ (g m−2 ) Point estimate Coefficient-ofvariation estimate 95% confidence interval 61.10 27.76 76.62 100 9.37 0.45 1597 211.54 227.86 22.37 0.13 0.11 0.77 0.01 0.17 0.13 0.28 0.17 0.35 0.35 (47.93, 75.54) (21.91, 32.38) (66.3, 500) (98.6,100) (5.97, 11.75) (0.20, 0.63) (1244, 3220) (149.57, 290.12) (224.71, 768.23) (15.34, 52.43) †Coefficients of variation were calculated by the non-parametric bootstrap (1000 replicates) with the transect as the sampling unit; confidence intervals were calculated by using the percentile method. researchers to examine the energetic value of krill swarms, which is important for estimating krill predator foraging strategies; this can be done by examining the energetic interconnectivity of swarms by applying percolation theory, for example (e.g. Reynolds et al. (2009)). Given that ambient levels of light are believed to trigger DVM (e.g. Zhou and Dorland (2004)), it may be possible to parameterize the vertical distribution function πy .y/ to include ambient levels of light at the surface as a parameter. However, this approach would require the likelihood to be rewritten and recent research suggests that DVM may not be a ubiquitous feature of krill behaviour. Tarling et al. (2009) found that ambient levels of light did not influence swarm depth in open ocean. Tarling et al. (2009) suggested that distance to land-based air breathing krill predator colonies may explain the magnitude of DVM, hypothesizing that krill outside predators’ foraging range are not required to seek refuge in deeper waters and so do not undertake DVM. There are many potential additional parameters for both the vertical distribution PDF and detection functions, but, as the above DVM example illustrates, they may not contribute information to the likelihood. As with other distance sampling methods, edge effects exist, but these are taken account of in estimation in two ways. Firstly, we model the reduction in the probability of detection as M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley 0 5 10 15 14 0.0 0.2 0.4 0.6 0.8 1.0 θ Fig. 7. Estimate of the mean krill swarm biomass: correcting for size-biased angular detection probability (the individual krill swarm biomass estimated by using the intercept of the regression of log(swarm biomass) on the detection angle (see the text for details)): - - - - - - - , 95% confidence bounds the deviation from vertical increases (see Fig. 5). This accounts for the reduced detectability of swarms towards the edges of the swath. Secondly, we model the dependence of observed biomass on the deviation from the vertical (see Fig. 7). This should account for a reduced average apparent size of swarms towards the edges of the pie slice (because swarms that are at greater angles from the vertical will tend to have greater fractions of their biomass falling outside the searched pie slice). Analyses of SBE data typically assume perfect detectability of krill swarms in the beam searched (e.g. Brierley et al. (1999) and Reiss et al. (2008), but see Demer et al. (1999)). In our analysis of MBE data we have not made this assumption; how different would estimates be if we had made the assumption? We estimate that the swarm detection probability has dropped by about 10% at 60 m and by about 20% at 100 m. The effect of this decline on density and biomass estimates depends on the vertical distribution of krill. To investigate this for our data, we reanalysed the survey data by assuming perfect detectability, i.e. p.r, θ/ = 1 for Å all (r, θ) in the swath searched (Fig. 2). In this case, the MLE of mean detection probability P̂ is higher by about 20% and estimates of abundance N̂ s , biomass B̂ and density ρ̂ are lower by about Å 12% (Table 4). The coefficient of variation of P̂ shrinks by a factor of about 3 when perfect detectability is assumed, but this translates into a reduction of less than 2% in the coefficient of variation of N̂ s , B̂ and ρ̂. If we assumed perfect detectability for folded MBE observation angles, θ > 0:7 rad, then N̂ s -bias would be less than 12%. Although these biases are relatively small, we recommend that the probability of detection is estimated when estimating abundance, biomass or density from MBE data to avoid the bias. Estimating Density of Antarctic Krill 15 Table 4. Model parameter and krill biomass variance estimates assuming certain detection† Parameter Å P̂ N̂ s B̂ (tonnes) ρ̂ (g m−2 ) Point Coefficient-ofestimate variation estimate 0.54 1344.02 283.60 19.35 0.04 0.26 0.23 0.23 95% confidence interval (0.49, 0.57) (1101, 1984) (201.99, 455.59) (13.79, 31.09) †Coefficients of variation were calculated by the non-parametric bootstrap (1000 replicates) with the transect as the sampling unit; confidence intervals were calculated by using the percentile method. Although we have used the distance sampling methods that were developed here to estimate krill density and biomass, the methods are completely general for MBE surveys and are applicable to MBE surveys of other species, including single animals. Acknowledgements We thank the Royal Society for funds enabling deployment of the MBE, Simrad USA, particularly J. Condiotty, for the loan of the SM20, and D. Needham and M. Patterson of Sea Technology Services for designing and constructing the MBE transducer mount. We are grateful to the US Antarctic Marine Living Resources Program; the Master, officers and crew of the RV Yuzhmorgeologiya and the personnel at the Cape Shirreff field station for logistical, equipment and ship support. MJC was supported by a UK Natural Environment Research Council doctoral studentship. This work was carried out in association with a National Science Foundation funded (grant 06-OPP-33939) investigation of the Livingston Island near-shore environment. A National Science Foundation grant was held by DAD and J. D. Warren of the School of Marine and Atmospheric Sciences, Stony Brook University, USA. We are grateful to E. Rexstad and T. Marques for both the provision of R code and helpful discussions of distance sampling theory for MBE surveys. We thank A. Jenkins, T. S. Sessions and M. van den Berg for skippering the near-shore boats and T. S. Sessions for providing invaluable expertise in the field. References Atkinson, A., Whitehouse, M. J., Priddle, J., Cripps, G. C. and Brandon, M. A. (2001) South Georgia, Antarctica: a productive, cold water, pelagic ecosystem. Mar. Ecol. Prog. Ser., 216, 279–308. Brierley, A. S., Watkins, J. L., Goss, C., Wilkinson, M. T. and Everson, I. (1999) Acoustic estimates of krill density at South Georgia, 1981 to 1998. CCAMLR Sci., 6, 47–57. Buckland, S. T., Anderson, D. R., Burnham, K. P., Laake, J. L., Borchers, D. L. and Thomas, L. (2001) Introduction to Distance Sampling. Oxford: Oxford University Press. Cox, M. J., Warren, J. D., Demer, D. A., Cutter, G. R. and Brierley, A. S. (2010) Three-dimensional observations of swarms of Antarctic krill (Euphausia superba) made using a multi-beam echosounder. Deep Sea Res. II, 57, 508–518. Cresswell, K., Tarling, G. and Burrows, M. (2007) Behaviour affects local-scale distributions of Antarctic krill around South Georgia. Mar. Ecol. Prog. Ser., 343, 193–206. Demer, D. A. (2004) An estimate of error for the CCAMLR 2000 survey estimate of krill biomass. Deep Sea Res. II, 51, 1237–1251. Demer, D. A. and Conti, S. G. (2005) New target-strength model indicates more krill in the Southern Ocean. ICES J. Mar. Sci., 62, 25–32. 16 M. J. Cox, D. L. Borchers, D. A. Demer, G. R. Cutter and A. S. Brierley Demer, D. A. and Hewitt, R. P. (1995) Bias in acoustic biomass estimates of Euphausia superba due to diel vertical migration. Deep Sea Res. I, 42, 455–475. Demer, D. A., Soule, M. A. and Hewitt, R. P. (1999) A multiple-frequency method for improved accuracy and precision of in situ target strength measurements. J. Acoust. Soc. Am., 105, 2359–2376. Fewster, R., Southwell, C., Borchers, D., Buckland, S. and Pople, A. (2008) The influence of animal mobility on the assumption of uniform distances in aerial line-transect surveys. Wldlf. Res., 35, 275–288. Gerlotto, F., Soria, M. and Freon, P. (1999) From two dimensions to three: the use of multi-beam sonar for a new approach in fisheries acoustics. Can. J. Fish. Aquat. Sci., 56, 6–12. Lascara, C. M., Hoffmann, E. E., Ross, R. M. and Quetin, L. B. (1999) Seasonal variability in the distribution of Antarctic krill, Euphausia superba, west of the Antarctic Peninsula. Deep Sea Res. I, 46, 951–984. MacLennan, D. N., Fernandes, P. G. and Dalen, J. (2002) A consistent approach to definitions and symbols in fisheries acoustics. ICES J. Mar. Sci., 59, 365–369. Marques, T. A., Buckland, S. T., Borchers, D. L., Tosh, D. and McDonald, R. A. (2010) Point transect sampling along linear features. Biometrics, to be published, doi 10.1111/j.1541-0420.2009.01381.x. Reiss, C. S., Cossio, A. M., Loeb, V. and Demer, D. A. (2008) Variations in the biomass of Antarctic krill (Euphausia superba) around the South Shetland Islands, 1996–2006. ICES J. Mar. Sci., 65, 497–508. Reynolds, A. M., Sword, G. A., Simpson, S. J. and Reynolds, D. R. (2009) Predator, percolation, insect outbreaks, and phase polyphenism. Curr. Biol., 19, 1–5. Tarling, G. A., Klevjer, T., Fielding, S., Watkins, J., Atkinson, A., Murphy, E., Korb, R., Whitehouse, M. and Leaper, R. (2009) Variability and predictability of antarctic krill swarm structure. Deep Sea Res. I, 56, 1994– 2012. Zhou, M. and Dorland, R. D. (2004) Aggregation and vertical migration behavior of Euphausia superba. Deep Sea Res. II, 51, 2219–2137.
© Copyright 2026 Paperzz