Estimating individual fish school biomass using digital omnidirectional sonars, applied to mackerel and herring

Abstract

Economic profitability and responsible fisheries are objectives of fishermen and fisheries managers. In purse seine fisheries, an accurate biomass estimate of the targeted school is crucial to accomplish this. For this study, omnidirectional fisheries sonar was used to estimate individual school biomass of Norwegian spring spawning herring (Clupea harengus) and Atlantic mackerel (Scomber scombrus). A sonar sampling design based on professional skipper’s experience provided detailed information on school dimensions and acoustic backscattering. Using calibrated digital sonar data, school volume and fish densities were obtained, and school biomass computed. A positive linear relation was found between the estimated sonar school biomass and purse seine catches for both species (r² = 0.92; residual standard error, RSE = 4.7 t). Large variability in volume backscattering coefficient and uncertainty in side-aspect target strength (⁠$TS$⁠) are the main sources of discrepancy between the estimates and the catch. Using a 4 dB (39%) weaker mean TS for mean side-aspect $TS$ than the normal mean dorsal aspect $TS$ was needed for optimizing the 1:1 relationship between sonar biomass estimate and catch. Accurate estimation of single school biomass can reduce the catch of unexpectedly large schools, leading to improvements in economic efficiency and reduced release of dead or dying fish.

Clupea harengus, fishing vessels, omnidirectional sonar, purse-seining, school biomass, Scomber scombrus

Introduction

About 20% of world industrial catches of fish are made using purse seine gear, targeting schooling pelagic species (Watson and Tidd, 2018). In modern purse seine fishing, accurate estimation of school biomass can reduce the release of unwanted catches (i.e. fish slipping) to fit vessel loading capacities, vessel quota restrictions, or other regulations (Tenningen et al., 2017). Accurate estimates also reduce occurrences of net bursting (Misund and Beltestad, 1995) and improve vessel economics through improved fish quality (Digre et al., 2016). Together with the economic consequences of wrongly estimating school biomass in the pre-catch phase, there is also uncertainty in unaccounted slipping mortality (Huse and Vold, 2010). Unaccounted fishing mortality in the capture process may also reduce the quality of the catch data used for stock assessment (Marçalo et al., 2019). The fishing industry and regulating authorities have a need for better tools for accurate estimation of school biomass before the catching process starts.

Omnidirectional fisheries sonars (hereinafter referred to as sonar) are multibeam acoustic systems designed for use in commercial fishing of pelagic species, both for long-range searching (5–8 km) and short-range pre-capture inspections at 200–400 m range. Modern sonars transmit from a cylindrical array a nearly omnidirectional acoustic pulse around the vessel and the received echoes can be monitored in electronically formed beams, configured as a 360° conical shell of beams (often 64) where the fan inclination angle below the sea surface can be varied electronically. The sonar can usually also be configured for alternating operation between horizontal and vertical modes, with a vertical 180° fan of beams that can be orientated in any direction (Vatnehol et al., 2017).

In recent years, critical issues have been resolved that facilitate the use of sonars for scientific purposes, such as increased dynamic range (to about 140 dB) and the development of software to efficiently process sonar data. In addition, documentation of the data file formats produced by sonars has become available (Simrad, 2004; Tang et al., 2008), and a manufacturer-independent file format has been defined (Macaulay and Peña, 2018). Calibration procedures for sonars have been established (Nishimori et al., 2009; Bernasconi, 2012; Vatnehol et al., 2015) and equations to compute the volume backscattering strength (⁠$S_v$_, dB re 1 ${\mathrm{m}}_{,}^{-1}$MacLennan et al., 2002) and target strength (⁠$TS$⁠, dB re 1 m²) from sonar data output have been published (Macaulay et al., 2016).

When acoustic beams are oriented in a horizontal manner, like here, the backscattering of fish becomes more complex than for vertical echo sounding. The mean TS for fish in dorsal aspect is well documented, but data are mostly lacking for side aspect [notable exceptions are Boswell and Wilson (2008); Pedersen et al. (2009), and Lee and Kang (2010)] and those which are available are generally not at the frequencies commonly used by sonars (15–50 kHz). Side aspect $TS$ is highly influenced by pitch, roll and in particular yaw angles of the fish relative to the incident acoustic beam (Cutter and Demer, 2007). This is very relevant for surveying situations, but maybe less in purse seining situations when each school is generally encircled.

In fisheries research, sonars have been used for multiple applications from school counting to complex segmentation methods for studying school behaviour (Hewitt et al., 1976; Misund, 1993; Nishimori et al., 2009; Stockwell, et al., 2013; Vatnehol et al., 2018). By combining sonar measurements of school morphometrics and fish density from echo sounders, Misund et al. (1992) proposed a method to compute single school biomass for abundance estimation and biomass measurements in pre-capture conditions during purse seining. The same approach was used to compare estimates of single school biomass of two pelagic species (Atlantic herring, Clupeas harengus and Atlantic mackerel, Scomber scombrus) with purse seine catches (Misund, 1993). Using digital data from sonar, Nishimori et al. (2009) computed school biomass by applying an echo integration method and obtained good agreement between the sonar estimates and the skipper’s estimates for one single school, although not verified by catch. Despite these efforts, no procedure for accurate biomass estimation of single schools has been proposed that uses only the sonar data.

In this work, we propose a method for improving the accuracy of single school biomass estimates for two important commercial species, Norwegian spring spawning (NSS) herring and Atlantic mackerel, using calibrated sonars. The estimates have been verified by capturing the measured schools using commercial purse-seining. New tools for processing sonar digital data are also described.

Material and methods

Data were collected from schools of NSS herring and Atlantic mackerel in the Norwegian and North Seas. Acoustic data were collected in periods between November 2012 and November 2017 from one research vessel and four purse seine/pelagic trawl fishing vessels (Table 1). Vessels were equipped with either the Simard SX90 (Simrad, 2013) or the Simrad SU90 (Simrad, 2015) sonar models. Both sonars were operated in the 20–30 kHz frequency range using either single frequency (CW; bandwidth at 450 m range was 167 Hz) or frequency modulated (FM; bandwidth 500 Hz) pulses. The SU90 sonar has a longer transducer than the SX90, and at 30 kHz, it generates acoustic beams with a nominal 5° vertical opening compared to 7° with the SX90. Both sonars generated 64 beams within the 360° angle of operation. The sonar settings during the data collected varied due to operational reasons (Table 1).

Table 1.

Summary of vessels and sonars used for the measurements.

Vessel	Year	Length (m)	Sonar	Frequency (kHz)	Pulse type	Region
G.O. Sars	2012	78	SX90	26	FM	Norwegian sea and fjords
	2013		SX90	26	CW	North Sea
Artus	2013	50	SX90	30	CW	Norwegian sea and fjords
Kings Bay	2014	78	SU90	30	CW	North Sea
	2017		SU90	26	FM	Norwegian sea
Eros	2015	78	SU90	30	FM	North Sea
	2016		SU90	20	FM	Norwegian sea
Brennholm	2016	75	SU90	30	FM	North Sea
Eros	2017	78	SU90	30	FM	Norwegian sea

The pulse type was either a single frequency (CW) or a frequency modulated pulse (FM).

The sonars were calibrated prior to the measurements according to published procedures (Macaulay et al., 2016) and the calibration parameters were applied during post-processing. Overall, the calibration accuracy was ±0.5 dB, adding about 10% uncertainty to the mean backscattering estimates.

Detailed sampling of individual schools was based on the normal procedure used by most fishermen during commercial purse seining. First, during searching at 3.8–5.9 ms⁻¹ vessel speed, schools were detected at long range (1500–2000 m) by the sonar using the 360° horizontal fan tilted between 4° and 7° below horizontal, with a pulse repetition frequency of about 0.5 Hz. A selected, suitable school that was judged by the skipper to be between 100 and 200 t was approached at an inspection speed of about 2.7 ms⁻¹, adjusting the tilt of the horizontal beams and reducing the sonar sampling range, with the aim of ensonifying the midsection of the school with the horizontal fan. This procedure will reduce the bias of measuring a non-uniform fish distribution within the sonar beam (Vatnehol and Handegard, 2018). At about 300 m range, the vessel course was adjusted in order to encircle the school, keeping the school at a distance of 100–200 m on the starboard side of the vessel. During this encircling, the bearing of the vertical beams was manually adjusted to sample the centre of the target school. After three or four circuits around the school, the fishing vessel then caught the whole school by purse seining. With a successful catch of the whole school, with no uncertainty of fish loss, either by escapement from the seine or by a hole in the net during the catch process, the total catch weight was obtained from the vessel’s fish weighing system. This operates by calculating the difference between the volume of cooled water inside the vessel’s holding tank before the pumping in the fish and the volume of water removed from the tanks after the pumping. This procedure or method gives an estimate of total fish volume, which by a species-specific factor is converted into weight, with an accuracy of about 5% (Pål Cato Reite, skipper FV “Eros”, pers. comm.). This estimate was ultimately verified by the weighing system in a factory when the catch was processed ashore.

A research quota of 600 t of the target species was available for each period of effort (i.e. a research survey) with the aim of capturing several schools of 100–200 t biomass. Opportunistic data were also collected from fishing vessels during their normal commercial fishing operations, providing data of larger schools.

For additional research on small schools and to increase the number of data points, an amount of fish was transferred into a cylindrical net pen of 13 m diameter and 5 m depth. The net pen was made from nylon monofilament of 0.9 mm thickness with 33.6 mm mesh size (knotted diamond) to reduce the acoustic backscattered echo at 20–30 kHz. The floating system on the net pen was a 125-mm diameter hydraulic tube, automatically filled with compressed air when launched from the vessel. The tilt of the sonar horizontal beams was adjusted to minimize the backscatter from this tube. After measuring the fish biomass inside the net pen with the sonar in the same manner as for a free swimming school, the fish were pumped onboard into a separate tank for weight estimation and biological sampling.

The acoustic data were processed using the LSSS (Large Scale Survey System, Korneliussen et al., 2016) software, version 2.5.1. The LSSS module, PROFOS (PRocessing system for Omnidirectional Fishery sOnarS), was used to isolate the proportion of data identified as a fish school and to output the morphological features and mean backscattering from the school at each ping.

Across-beam school dimensions can be distorted and overestimated (smearing effect) when observed with sonar due to overlap between adjacent beams, where the smearing depends on the width of the beams relative to the between beam angle (Reid, 2000; Vatnehol, et al., 2017). For the SX90 and SU90 sonars used in this study, the beam width is ∼10° and the between beam angle is 5.6°, implying significant overlap between beams. Contribution from side lobes to the smearing effect is negligible for these sonars. To compensate for this smearing effect, a two-stage adaptive segmentation method (Holmin and Peña, in prep.) implemented in PROFOS was applied to isolate schools from the background noise (Figure 1). In the first stage of the adaptive segmentation method (Peña et al., 2013), the data are thresholded at a fixed value 10 dB above$S_{v,B}$ to produce an initial segmentation mask [the mask is grown by a flood-fill algorithm (Torbert, 2016) starting from an automatic or user-defined seed point]. The background noise is estimated directly from the data by applying a moving median filter of width 21 along range bins in each beam, followed by the median across all beams and across five pings for each range from the sonar. This results in a vector of the same length as one beam for each ping, and finally a moving Gaussian filter with a standard deviation of 100 samples along the result. The resulting vector of estimated noise was copied to all beams.

Figure 1.

Raw data from a herring school at 300 m to the starboard side of the vessel, horizontal (left panel) and vertical beams (right panel). Vessel track shows the school has been encircled three times. The segmented school has a white overlay and displayed the centre of mass in each ping (red dots). The red square indicates the geographical extension of the school over all the pings where it was segmented. In the vertical beams, the school is about 100 m deep. The colour scale represents the calibrated volume backscattering strength, Sv (dB).

The second stage defines a final threshold (⁠$FT$⁠, dB) compensating for the smearing effect, given by:

$$FT=S_{v,90}\mathrm{ }–\mathrm{ }\left(4.4+0.013\mathrm{ }S+0.1\mathrm{ }\mathit{SNR}\right),$$

(1)

where $S_{v,90}$ is the 90-percentile of the $S_v$ inside the initial segmentation mask serving as a robust estimate of the characteristic backscatter from the school, $S$ is the horizontal size (diameter of a circle of the same horizontal area) of the school estimated from the initial segmentation mask, and $\mathit{SNR}$ is the signal-to-noise ratio defined by the ratio between mean $S_v$ and mean $S_{v,B}$ inside the initial segmentation mask. This equation was derived from an experiment where sonar data of a 12-h mini survey were simulated (Holmin et al., 2016) from 120 uniformly distributed and polarized schools of herring with variable ellipsoidal shape, and size between 16 and 330 m. For each simulated school, the threshold needed to obtain an estimated horizontal area equal to the modelled horizontal area was regressed linearly against $S$ and $\mathit{SNR}$⁠, resulting in Equation (1). Only the ten pings of largest estimated area of each school were included in the regression to avoid overrepresentation by schools with a high number of detections (i.e. at a favourable distance from the vessel). The performance of the adaptive segmentation method was verified for sonar data of a submerged net pen (diameter 13 m) and the cylindrical concrete legs of an oil platform (diameter 160 m, Troll A platform in the North Sea) at different distances. The method was found to be only minorly affected by changes in the initial threshold, as long as the initial segmentation mask was not dominated by noise samples.

The segmented school candidate is later evaluated using the heterogeneity of its cells. School candidates with homogeneous $S_v$ values are more likely to be noise from surface or bottom echoes. A school candidate in a particular ping was rejected if there were no sonar samples in a segmented school with an $S_v$ higher than the relative threshold plus an empirically defined additional 2 dB.

Once the selected school is segmented, PROFOS searches for ten adjacent pings, before and after the ping where the school was seeded, for the geographically corrected position of the seeded school, to continue with school segmentation. This procedure repeats until the size and the backscattering from the school is reduced by about 20%, being an empirically defined allowance to retain only the strong core of the school.

The school segmentation produces a label for each detected school (Figure 2) with associated ping-based information from schools segmented in the horizontal and vertical beams (Table 2).

Figure 2.

Sonar processing workflow steps starting with identification of a candidate school (top) and ending with an accepted and delineated school (bottom). The sonar echograms provide a visual illustration of the processing progress.

Table 2.

Ping-based information from schools segmented from horizontal and vertical beams.

Variable	Units	Horizontal beams	Vertical beams
Time stamp		x	x
Geographical position of the school centre of mass		x	x
School mean depth	m	x	x
Mean school Sv	dB	x	x
School area	m2	x	x
Maximum school crosswise extent	m	x
Maximum school lengthwise extent	m	x
Vessel geographical position		x
Vessel speed	ms−1	x
Sonar tilt angle	deg	x
School height	m		x
School width	m		x

Exploratory analysis of the computed school parameters showed that mackerel school No. 9 had anomalous low $S_v$ values and only limited data were collected from mackerel school No. 10 during the inspection phase. Therefore, both schools were removed from further analysis.

Sonar data collected during repeated circuits tend to show high variability in mean and total backscattering from the detected school (Holmin et al., 2016), depending on which direction the school was ensonified from relative to its actual swimming direction. For most schools, the backscattering is stronger in the side aspects (i.e. along-track), compared to when the schools are ensonified from the head or tail directions (i.e. across track) and is an indication of the degree of polarization of the fish inside the school (Holmin et al., 2016). For these reasons only data from along-track was used, as it is less influenced by background noise (Holmin, 2012; Takahashi et al., 2016). We also assume that all fish in the school are similarly oriented, a reasonable assumption for a schooling fish (Pitcher and Parrish, 1993).

We assume that the school shape can be described as an oblate spheroid, where the school area from the horizontal beams (⁠$a_h$⁠) is equated to the area of the sagittal section of the spheroid and the vertical school extension (⁠$b$⁠) obtained from the vertical beams provides the spheroid minor axis. The school volume, $V$⁠, (m³) is then:

$$V=\frac{4}{3}{a}_h\frac{b}{2}.$$

(2)

Fish density (⁠${\rho }_v$⁠, m⁻³) can be estimated using the volume backscattering as:

$${\rho }_v=\frac{s_v}{{\sigma }_{bs}},$$

(3)

where $s_v$ is the school volume backscattering coefficient (m⁻¹) and ${\sigma }_{bs}$ is the mean backscattering cross-section of individual fish (m²). PROFOS calculates a mean volume backscattering strength (⁠$S_v$ dB re 1 m⁻¹), being the decibel form of the average of all $s_v$ values within the boundaries of a school for each ping.

The TS (dB re 1 m²) of a fish is the logarithmic expression of the backscattering cross-section (⁠${\sigma }_{bs}$⁠) and reflects the acoustic reflectivity of an ensonified target. For computing the school biomass, and in absence of in situ side aspect TS, we use a modification of the equations for herring (Ona, 2003) and mackerel (ICES, 2005) for dorsal aspect TS at 38 kHz. For Atlantic herring this is:

$$TS=20\hspace{0.17em}\log \left(L\right)-2.3\hspace{0.17em}\log \left(1+\frac{z}{10}\right)-65+\mathit{Corr},$$

(4)

and for Atlantic mackerel is:

$$TS=20\hspace{0.17em}\log \left(L\right)-86+\mathit{Corr},$$

(5)

where $L$is fish length (cm), $z$ is fish depth (m), and $\mathit{Corr}$ is the correction factor required to convert the dorsal into side aspect TS. Fish length was obtained from the same sample used to obtain the mean weight.

Because of the use of dorsal $TS$⁠, a slope different from unity is expected when regressing the biomass derived from the sonar measurements and the purse seine catch (⁠$\mathit{Catch}=\alpha \mathrm{ }+\beta \mathit{Biomass}$⁠). An optimization method was used to find the value that would result in a slope, $\beta$⁠, equal to 1 and with an intercept, $\alpha$ of 0. The logarithm of this value is $\mathit{Corr}$ in Equations (4) and (5). This approach assumes that other parameters in the equation are not biased.

The school biomass $B$ (t) is then defined as:

$$B=\frac{{\rho }_{v}V\overline{W}}{1000},$$

(6)

is the mean weight of the fish (kg). Mean fish weight was estimated from fish caught by pelagic trawling and purse seining by R/V “G. O. Sars” and chartered fishing vessels, respectively, using a standard sample size of 100 fish.

Since the sonar alternates between horizontal and vertical ensonification, both horizontal school area $a_h$ and vertical school extension $b$ cannot be present at one ping. Therefore, pairs of horizontal and vertical pings separated by at most 4 s in time were combined to represent one sampling unit. For consecutive pings separated by more than 4 s, the information from either the vertical or horizontal fan was considered missing. In general, data from either the horizontal or vertical fan could be missing, e.g. if the school at certain pings was too weak to be segmented in the horizontal fan. To enable estimates at each sampling unit, missing data were imputed using the bootstrap expectation–maximization algorithm provided by the function amelia of the R package “Amelia” (Honaker et al., 2011). The algorithm bootstraps the data to estimate the distribution of the missing data conditional on the present data and draws from that distribution to impute the missing data. The variables included in the imputation were $s_v$⁠, area, and school extent from the vertical and horizontal fan. The advantage of this imputation method is that the variability of the data is preserved also in the imputed data, as opposed to simpler approaches such as estimating missing values by averaging the present data from neighbouring pings.

Results

A total of 76 purse seine sets were made within an effort of about 200 ship days over 9 fishing or research cruises. Single school biomass was able to be estimated from 15 of these, about 20% of the total (Table 3). About 49% of the sets resulted in no catch because of fish escapement, net breakage, or vessel mechanical problems during the catch process. The remaining 31% of the sets resulted in a catch but were not used because some of the targeted school escaped (21%) or the sonar data were inadequate due to incomplete school coverage (10%).

Table 3.

Summary of schools, catch information, and relevant biological fish properties.

School ID	Date	Time (UTC)	Vessel	Catch (t)	Species	Mean fish weight (g)	Mean fish length (cm)	Target
1	17 November 2013	06:41	Artus	110	Herring	340	35	School
2	12 November 2013	11:17	G.O. Sars	30	Herring	340	35	Net pen
3	29 October 2014	18:05	Kings Bay	6	Mackerel	320	33	Net pen
4	21 October 2015	16:34	Eros	75	Mackerel	369	35	School
5	24 October 2015	15:26	Eros	165	Mackerel	360	34	School
6	30 October 2015	08:58	Eros	203	Mackerel	340	34	School
7	22 September 2016	07:10	Brennholm	195	Mackerel	377	35	School
8	1 October 2016	15:05	Brennholm	251	Mackerel	382	35	School
9*	29 September 2016	12:29	Eros	540	Mackerel	395	36	School
10*	12 October 2017	10:48	Kings Bay	607	Mackerel	430	37	School
11	7 November 2017	20:02	Eros	94	Herring	371	35	School
12	13 November 2017	16:24	Eros	100	Herring	372	35	School
13	13 November 2017	19:03	Eros	43	Herring	379	35	School
14	18 November 2017	22:11	Eros	85	Herring	347	34	School
15	19 November 2017	16:27	Eros	125	Herring	378	36	School

Schools from commercial fishing operations and not related to a research quota are noted with an asterisk.

Purse seine catches ranged from 43 to 607 t with an average catch 198 t. Caught herring had a mean weight of 362 g (SD = 16.2 g) and a total length of 35 cm (SD = 0.6 cm). Mackerel had a mean weight and total length of 372 g (SD = 31.4 g) and 35 cm (SD = 1.2 cm) (Table 3). During the inspection phase, when encircling schools for sonar sampling, systematic changes in volume backscattering strength (⁠$S_v$⁠) were observed. Mackerel schools in general showed a higher degree of polarization, with changes of up to 15 dB (Figure 3), which based on school direction and speed correspond to whether the schools were ensonified from the side (higher backscattering) or from the head or tail (lower backscattering). Herring schools had a lower level of polarization, and less variability in mean school $S_v$ was observed during the encircling. Some schools also became smaller and more compact after being encircled repeatedly by the vessel, indicating a reaction of the fish to the vessel, and potentially acoustic extinction (Furusawa et al., 1992). Data from the initial part of the encircling were therefore preferred.

Figure 3.

Vessel and mackerel school No. 6 positions during a 7-min inspection period prior to setting a purse seine (left panel). Start and end of vessel and school tracks are noted with a square and triangle, respectively. Mean school Sv measured by the sonar aggregated in 20-s intervals (right panel). The vessel and school positions when maximum Sv are indicated by continuous (labelled 1) and dashed (labelled 2) lines in left panel. Maximum Sv values (ca. −50 dB) when school was ensonified from the side (right panel). Minimum values were measured when school was measured from head and tail aspect.

Although data for biomass estimation were selected from periods of maximum cross-sectional school area, across vessel track, school parameters showed some variation between pings (Figure 4). The dispersion of school parameters for both species (lengthwise extent, area, and height) fluctuated with a median coefficient of variation (CV) of between 11 and 22% (Figure 5). The largest variation was observed in the volume backscattering coefficient, with a median CV of 43%, while the other parameters had a lower CV (30%).

Figure 4.

Properties of mackerel school No.5, given as an example of changes in school parameters per ping during the measurement period (about 50 min).

Figure 5.

Coefficient of variation (CV) for school parameters: volume backscattering coefficient (sv), lengthwise extent, area, and height.

By species, free-swimming herring schools showed much higher backscattering, with mean S_v of −46 dB (25th percentile: −49 dB, 75th percentile = −44 dB) compared to free-swimming mackerel schools with a mean $S_v$ of −54 dB (25th percentile: −56 dB, 75th percentile = −52 dB) (Figure 6).

Figure 6.

School volume backscattering strength (left panel), school volume (centre panel), and fish density (right panel) for mackerel and herring derived from the experiments. Data from net pens (schools No. 2 and 3) were excluded because of unrealistic parameters in comparison to free swimming schools.

Data from net pens (schools No. 2 and 3) were excluded from the density analysis because of the unrealistically high packing density compared to free swimming schools. Mackerel schools contained higher densities with a median of about 3.4 fish m⁻³ rising to a maximum of 20.6 fish m⁻³ (25th percentile: 1.9 fish m⁻³, 75th percentile = 5.4 fish m⁻³) (Figure 6). Herring densities were typically about 0.3 fish m⁻³ with a maximum of 3.4 fish m⁻³ (25th percentile: 0.2 fish m⁻³, 75th percentile = 0.6 fish m⁻³).

The catches from research surveys were below 250 t due to the restriction imposed by the limited quota available and the need to maximize the number of purse seine catches in each survey (Table 4).

Table 4.

School statistics: school volume (m³), fish density (fish m⁻³), sonar school biomass estimate (t), and purse seine catch (t).

School Id	Volume (1st Qu.)	Volume median	Volume (3rd Qu.)	Density (1st Qu.)	Density median	Density (3rd Qu.)	Biomass (1st Qu.)	Biomass median	Biomass (3rd Qu.)	Catch
1	187 248	282 168	362 862	1.4	2.0	2.6	89	130	148	110
2	2993	4044	5747	2.6	6.9	13.5	5	8	15	30
3	881	1247	1690	10.7	21.8	36.0	4	8	14	6
4	62 355	76 207	88 571	2.7	4.4	5.4	50	106	146	75
5	39 563	48 658	67 510	5.6	9.2	14.6	88	138	206	165
6	119 851	137 265	160 251	4.0	6.1	8.2	179	245	290	203
7	186 369	219 852	267 140	0.9	1.8	2.8	49	127	194	195
8	224 489	273 898	326 820	1.4	2.5	3.6	104	215	336	251
11	611 974	717 583	916 584	0.2	0.3	0.27	42	59	75	94
12	344 639	419 457	496 566	0.5	0. 7	0.9	53	77	128	100
13	510 352	555 275	604 502	0.1	0.2	0.2	23	28	33	43
14	1 096 579	1 448 966	1 985 118	0.2	0.2	0.3	82	94	104	85
15	1 383 907	1 567 727	1 772 063	0.2	0.4	0.8	93	199	314	125

Schools No. 9 and 10 were removed from the analysis due to poor data quality and are not listed here.

The optimization method yielded a value for $\mathit{Corr}$ of −4.5 dB for herring and −4.1 dB for mackerel. A very strong positive linear relation between the estimated sonar school biomass and the purse seine catches (obtained from the factory’s fish weighing system) was found for both species (r² = 0.92; residual standard error = 4.7 t) (Figure 7). The prediction intervals increased with the larger catches. From these results, for a computed sonar biomass of 200 t, a catch of 200 t is predicted with a 95% prediction interval between 161 and 239 t. By species, herring showed a strong positive relation (r² = 0.89; RSE = 5.1 t), similar to mackerel schools (r² = 0.94; RSE = 4.6 t) (Figure 7).

Figure 7.

Sonar biomass estimates and purse seine catch for all mackerel and herring schools (upper panel). Solid line shows the predicted relation (r² = 0.92; RSE = 4.7 t) and the grey polygon the prediction interval weighted by the inverse of the standard deviation. Bottom row shows data for herring (left panel, r² = 0.89; RSE = 5.1 t) and mackerel (right panel, r² = 0.94; RSE = 4.6 t). Numbers indicate school unique identificatory number.

Discussion

The relevance of the present work resides in the absence of accurate means to estimate fish school biomass before targeting for commercial purse seining. Skippers rely mostly on their experience and mental reconstruction of the school volume to assess the school biomass. However, when school densities vary, large unwanted catches can occur with this method. A too large catch may have destructive effect on the gear; thus, a more accurate method is needed, and one based on sonar measurements is preferable. We note that echo sounder measurements are adversely affected by school reactions when overrun by a vessel at depth ranges suitable for purse seining (i.e. shallower than 100 m).

School volume overestimation will result in larger biomass estimations when used to compute the total number of fishes in a school. Correction factors for distortion and overestimation of school dimension caused by the beam pattern effect have been proposed previously (Misund, 1990a; Reid, 2000; Vatnehol et al., 2017; Trygonis and Kapelonis, 2018). However, these corrections were made for shorter ranges or did not consider the large beam overlap that was contained in the sonar data. Our correction method was applied to both horizontal and vertical beams, where the school area and height were used to compute the school volume. In previous work, school area estimates had a large variability when compared to similar catch sizes, assuming similar fish densities. Misund (1993) presented a relation between school area measured with sonar and catches, i.e. for a 200 t school the area computed was 4 634 m². Misund et al. (1992) used vertical echo sounder echo integration to calculate a biomass from 1 to 75 t for the same school area. The large difference and unrealistic results were explained by the sampling method or the assumption of a circular school area. For Japanese mackerel, Tang et al. (2008) used digital sonar data and similar area corrections used by Misund to compute larger area estimates of 1000 and 100 000 m² for catches of 10 and 65 t, respectively. From our study, for a herring school of 203 t, the measured area was close to double that reported by Misund (1993) with 8500 m². We consider that our results are more accurate because the cross-sectional area of the school is the sum of the acoustic samples area contained in the school, with no ideal shape assumption, and correction for the beam pattern effect.

Vertical school extent in Misund’s method (1990a) is derived from echo sounder measurements, assuming an absence or low degree of disturbance in fish behaviour when the school is approached and overrun by the surveying vessel. In addition, the time mismatch between the sonar and echo sounder measurements, as they were not recorded at the same time, was disregarded. From surveys in 2012 and 2013, we found large variability in vertical school extent and acoustic volume density from echo sounder measurements during repeated sampling of the same school, reinforcing the need to develop a non-intrusive method for biomass estimation of individual schools. This is consistent with the review by De Robertis and Handegard (2013).

Herring density in free swimming schools varied between 0.1 and 0.9 fish m⁻³ (first and third quantiles, Table 4). These values are higher with the 0.05–0.13 fish m⁻³ range reported elsewhere (Nishimori et al., 2009) using sonar data to compute the school volume and fish volume density estimates from an echo integration approach. Higher densities have been reported using manual measurements of projected school dimensions and echo sounder fish densities estimates; these results included values of 4.8 (SD = 3.1) fish m⁻³ (Misund and Øvredal, 1988), 1–2 fish m⁻³ (Misund, 1990b), an average of 4.3 fish m⁻³ (0.3–22 fish m⁻³; Misund and Aglen, 1993) and between 0.7 and 7.3 fish m⁻³ (Misund and Beltestad, 1995). Large differences could be related to the different methods for estimating school volume and acoustic volume densities, and differences in fish behaviour and environmental conditions. Mackerel densities of between 1.1 and 8.1 fish m⁻³ (Misund and Beltestad, 1995) are similar to our findings of 0.9–14.6 fish m⁻³ (first and third quantiles).

Beam incidence angle relative to the fish orientation within a fish school is an important parameter when estimating side aspect TS for sonar applications (Frouzova et al., 2005; Cutter and Demer, 2007; Pedersen at al., 2009, Holmin et al., 2012, Takahashi et al., 2016). Herring models indicate that $TS$ can vary by up to 20 dB in yaw angles from 0 to 90° (Pedersen et al., 2016). Similar results were found with modelled and in situ data (Cutter and Demer, 2007), with a decrease in normalized $TS$ (by the maximum extent of the model surface by aspect), from a value of 0.9 when fish were oriented along-track to 0.4 when fish were oriented across-track. We found similar results in school $S_v$⁠, where differences of between 4 and 10 dB were observed when polarized schools were ensonified between along or across-track (Figures 3). Even though the adopted criterion was to use sonar data aiming for along-track ensonification, (higher $S_v$ and larger area), the variability in $S_v$ was larger than expected and the larger of the critical parameters used (Figure 5). Variability could be explained by the inclusion of non along-track sonar data, which could reduce backscattering by about 30% when fish is ensonified at 30° away from along-track (Cutter and Demer, 2007). Larger differences were observed in herring schools measured by Tenningen et al. (2017) of about 10 dB lower when a school was ensonified with an angle of 20° off the centre of the school. Ensonification angles computed from school tracking and vessel displacement at regular intervals could be used to refine the selection of data for biomass estimation, as suggested by Tang (2004).

Because of the lack of in situ side aspect TS measurements, the well-known dorsal TS, reported at 38 kHz was initially used to compute the school biomass for both herring and mackerel, while waiting for accurate side aspect TS data. The closest echo sounder frequencies are 18 and 38 kHz, and the expected $TS$ in this frequency region may be extracted from models or from practical multifrequency work (Korneliussen, 2010). Using the difference in r(f) from many surveys and experiments, Korneliussen concluded that the mean backscattering from adult mackerel is 1.2 (20%) times stronger at 18 kHz than at 38 kHz. Expected, interpolated TS at 28 kHz (middle sonar band) is less than this (10%), or 1.3 dB stronger than for 38 kHz. Furthermore, in situ measurements of mean side aspect TS at the available frequencies 38, 70, 120, and 200 kHz (EO, pers. comm.) show a mean TS from side aspect to be about 4 dB weaker than dorsal aspect both for adult herring and mackerel, the difference increasing with frequency, 38–200 kHz. If interpolated to 28 kHz, the expected difference is then about 2.7–3.0 dB weaker than the dorsal $TS$ for mackerel at 38 kHz.

The $\mathit{Corr}$ values obtained from the optimization method confirmed that a lower $TS$ was required to obtain a 1:1 relation between the biomass estimates and the catch. The reduction of about 4 dB was larger than the 1–2 dB indicated by modelling of herring at 50 m depth (Pedersen et al., 2009). These results differ from measurements at 38 kHz on herring at 50 m depth (Pedersen et al., 2016), when an increase of 3 dB was reported. Models of side aspect backscattering also show higher $TS$ than dorsal aspect (Cutter and Demer, 2007; Tang et al., 2009). These discrepancies are related to the fish yaw angle distribution used when averaging during the modelling, but also to the variable angle distribution of free-swimming fish during in situ measurements compared to ex situ measurement. In addition, the range of ensonification angles of the sonar beams in every school detection was about 40°, which increases non-normal ensonification. These differences highlight the need for in situ side-aspect TS at sonar frequencies.

The range of the catches was limited by the research quota assigned during each survey, and therefore underrepresent the more frequent large school sizes (>150 t) captured by the Norwegian purse seine fleet (Data from Norwegian Directorate of Fisheries, www.fiskeridir.no). For larger schools than the ones presented, methods for estimation of, and corrections for acoustic extinction are needed for swimbladdered fish.

At present, some fishery sonars provide a real-time school biomass estimate based on school cross-sectional area or volume and a constant fish area density (i.e. 25 kg m⁻²). No data were available to compare biomass estimates from commercial sonars and our results, which would be advisable for future studies. Sonars used can compute a biomass estimate using a target tracking method by automatic tilt and bearing. This method was not the appropriate for our study where manual tilt and bearing was required.

Critical aspects for school biomass estimation were approached and parametrized. Without the means to compare the sonar estimates with ground truth from real catches the present work would be just a theoretical exercise. Therefore, linking catches to sonar measurements provides a valuable data set. In previous studies, catch data from single schools was obtained from commercial fishing operations with no control of the targeted schools nor the sampling design (Misund, 1990a; Misund 1993; Nishimori et al., 2009).

During normal purse seining, in most cases, skippers attempt to catch part of a layer or large school, or two schools if they are small with a consequent risk of fish escapement. The criteria used during the work reported here improved the probability of capturing single schools and minimized fish escaping under the net or below the vessel during the net pursing. The low number of valid school data points (13) is a result of the difficulty in meeting all these criteria, combined with the costs involved (vessel and fish quota). An alternative to increase the number of data points was to join commercial fishing trips, which eventually can provide access to more and larger school sizes. However, the number of complete catches of a single school was low, because of not targeting catch a whole school, instead a fraction of a large school or layer. Also, as observed with school No. 10, some skippers did not encircle the school, or weather and fishing conditions prevented encircling, resulting in high uncertainty of the school size and density. In these cases, the risk of obtaining larges catches is higher, with the consequent need to release part of the catch.

Summary

Our work presents an updated and improved methodology for single school biomass estimations for herring and mackerel based on digital calibrated sonar data. A very strong positive linear relationship was observed between the estimated sonar school biomass and the purse seine catches for both species. Some questions remain with respect to which mean TS should be used for the conversion of mean volume scattering strength to biomass. Here we have used the catch data to obtain a correction factor for the dorsal TS used. We expect that our model is better than the ones applied in fishery sonars today, due to the removal of effects from acoustic beam smearing and by accounting for differences in packing density between schools. Also, due to the acoustic directivity of the single targets, and hence the whole school when moving, there is a fair uncertainty in the biomass estimate prior to shooting the net if the school is not encircled and measured intensively with the sonar. Skipper’s catch procedure and fish behaviour play a decisive role in the success of catching the desired size of school and the methodology presented here may reduce unwanted fish mortality in the catching process of a limited, specific fish quota.

Data availability

All data and results are presented in the manuscript. Provision of raw data will be considered upon request.

Acknowledgements

The cooperation and enthusiasm of the skippers and crew of research vessel “G.O. Sars”, and fishing vessels “Artus”, “Brennholm”, “Eros”, and “Kings Bay” is gratefully acknowledged. Atle Totland’s valuable contribution during the sonar calibration and early stages of this study is acknowledged. Dr Maria Tenningen is thanked for collecting sonar data during opportunistic fishing.

Peña, H., Macaulay G. J., Ona, E., Vatnehol, S., and Holmin, A. J. 2020. Estimating individual fish school biomass using digital omnidirectional sonars, applied to mackerel and herring. – ICES Journal of Marine Science, 00: 000–000.

Funding

This work was funded by the Norwegian Research Council through Grant Numbers 216460 (WHOFISH Whale counting and fish school biomass appraisal by two new omni-directional fishery sonars) and 203477 (CRISP Centre for Research-based Innovation in Sustainable Fish Capture and Processing Technology).