Br oadband backscat tering by Atlantic her r ing ( Clupea harengus L.) differs when measured from a resear c h v essel vs. a silent uncrewed surface vehicle

Broadband frequency-modulated signals are believed to improve acoustic spectral-based target classification. Efficient use of uncrewed surface vehicles (USV) for fisheries science applications, with no possibility for biological sampling, is believed to be facilitated by use of broadband signals with methods for target classification. If the broadband frequency response used to train automated target classifiers are obtained from conventional research vessels (RVs), due to potential vessel avoidance, the swimming angle distribution may be different than for USVs. This may have consequences for target classification if the model is trained with RV data. The aim of this study was to assess whether the frequency response differs between platforms due to avoidance. Broadband acoustic data were collected with a conventional RV and a small USV. The broadband frequency response of Norwegian spring spawning herring obtained with the USV and RV was found to be significantly different for shallow herring layers in the 200 kHz band. This indicates that broadband frequency response has potential as a tool for real-time monitoring of behaviour reactions to vessels and to provide insight into fish behaviour in general. When using broadband frequency response for target classification, the potential platform-dependent broadband frequency response should be considered.


Introduction
Fish population surveys that use echosounders provide valuable data about the distribution and abundance of fish, which is crucial for evaluating and managing fisheries (MacLennan and Simmonds 2005 ).Acoustic target classification (ATC) (Korneliussen 2018 ) is the process of using the acoustic signal and auxiliary data to identify and classify acoustic backscatter to species or groups of species, known as acoustic categories.Different methods can be used to gather and analyse the data from acoustic surveys, but traditionally, the information from the echosounder is combined with additional data from trawl sampling in a manual classification process.Traditionally, acoustic surveys use echosounders with discrete frequencies, but the latest broadband echosounders are able to generate a nearly continuous frequency spectrum by combining the output from multiple transducers (Demer et al. 2017 ).The increased frequency resolution of the target is suggested to improve spectral-based target classification (Stanton et al. 2010 ).
Automated methods for ATC are based on a range of different algorithms (Korneliussen 2018 ).Machine learning algorithms applied to discrete multi-frequency data include methods that rely on manual feature extraction, e.g.Haralabous and Georgakarakos 1996 ; traditional region of interest detections passed on to deep learning classifiers (Rezvanifar et al. 2019 ); and supervised deep learning methods based on convolutional neural networks (Brautaset et al. 2020 ), among others.There are several different methods that use broadband data to analyse acoustic survey data, including the use of random forest to distinguish sticklebacks and whitefish in a controlled environment (Gugele et al. 2021 ), applying unsupervised clustering to individual targets and comparing the results to known models (Agersted et al. 2021 ), and training a k -Nearest Neighbors algorithm using scattering data from five different target types (Cotter et al. 2021 ).
Another recent technological development is the increased availability of platforms capable of carrying acoustic sensors.Acoustic sensors have been deployed on, e.g.AUVs (Fernandes et al. 2003 ), gliders (Davis et al. 2008 ), and uncrewed surface vehicles (USVs), with extended persistence powered by renewable energy sources such as wave gliders (Greene et al. 2014 ), sailing drones (Hauge et al. 2016, Mordy et al. 2017, De Robertis et al. 2021 ), and motorized USVs (Totland and Johnsen 2022 ).These platforms have a considerably lower running cost than research vessels and will lead to an upscaling of acoustic data collection.Combined with automated data processing and analysis methods (Malde et al. 2020 ), this may result in more information of the marine ecosystem compared to the traditional ship-based data collection and manual data processing steps.
Overwintering Norwegian spring spawning (NSS) herring ( Clupea harengus ) in Norwegian fjords generally form large and dense layers (Huse and Ona 1996 ).These layers are found at shallow depths during the night and deeper during the day.During the overwintering, the fish are mainly avoiding predation and conserving energy, waiting for the southward spawning migration along the Norwegian coast (Huse and Ona 1996 ).There are several predators in the area, including saithe and killer whales (Nøttestad 1998 ), and the fish actively respond to whale attacks (Nøttestad and Axelsen 1999 ).
An approaching research vessel (RV) may affect the fish behaviour (De Robertis and Handegard 2013 ), either by displacing the fish horizontally (Soria et al. 1996 ), provoking a diving response (Vabø et al. 2002 ), or both.It is also expected that the behaviour is different in response to smaller platforms (Stoner et al. 2008 ), like AUVs (Patel et al. 2004 ), gliders (Guihen et al. 2022 ), and USVs (Totland and Johnsen 2022 ), as well as differences between vessels of different sizes and noise characteristics (Ona et al. 2007 ).For RVs in general, fish < 200 m depths are mostly unaffected; fish between 100 and 200 m depth may respond, whereas at depths shallower than 100 m, there are typically substantial avoidance reactions (De Robertis and Handegard 2013 ).It is also worth noting that even if there is a significant diving reaction, the integrated backscatter may (or may not) be similar to the undisturbed case (Ona et al. 2007 ).
Broadband acoustic echosounders are becoming more widely used in fisheries acoustics, where frequency-modulated (FM) pulses are transmitted either sequentially or simultaneously, potentially spanning a more or less continuous spectrum from 10 to 500 kHz (Demer et al. 2017 ).In addition to the improved range resolution, potentially improved signal-to-noise ratio, and target identification, a continuous spectrum may reveal behavioural cues that are not visible when using discrete narrow-band frequencies.Relative frequency response, as commonly employed with narrow-band echosounders (Korneliussen et al. 2009, 2016, De Robertis et al. 2010 ), has been employed on broadband data, e.g. by Bassett et al. (2018) for discrimination between fish and euphausiids.The use of relative frequency response removes amplitude offsets, which are caused by, e.g.density of organisms in an aggregation, while preserving spectral information connected with behaviour and morphology/physiology.Broadband has previously been explored for direct extraction of information on fish orientation, but to our knowledge not using the frequency response specifically.Some examples of studies include Tušer et al. (2023) , which studied the echo envelope of tethered fish, and Stanton et al. ( 2003 ), which examined the potential in discerning internal scattering features.
Changes in behaviour may change abundance estimates by changing the S v (volume backscattering strength, dB re 1 m −1 ), and consequently the S v based frequency response, and affecting ATC performance.Our main focus in this paper is on the relative frequency response using broadband S v .Where the relative frequency response is defined as s v (volume backscattering coefficient, m −1 ), the linear version of S v , for a specific frequency relative to that of a reference frequency.If the frequency responses used to train the automated ATC are obtained from the vessels, which affect fish behaviour, the difference in tilt angle distribution of the fish may lead to a biassed performance of the automated ATC.Applying the classifier trained on one situation to the other may lead to weaker automated ATC performance.Another approach is to use model-based predictions to train the ATC models (Cotter et al. 2021 ), but then behavioural parameters like tilt angle distributions are required for reliable predictions.This may cause challenges for target identification, and no matter the approach, the data must be representative for the field data that may be platform-dependent.On the other hand, if the broadband frequency response changes are predictable with respect to behaviour, e.g. through modelling different tilt angle distributions, the patterns in the frequency response may be used to indicate whether there are avoidance behaviours during acoustic data collection for a given platform or not.
The aim of this study is to investigate the broadband frequency response of fish when approached by a RV and a USV.Based on earlier work, we hypothesize that the behaviour of the fish will change in response to the presence of the vessel but not for the USV, and that any change in behaviour will cause the frequency response to change.Avoidance, or lack thereof, is documented using a moored inverted echosounder, and frequency responses obtained with a traditional research vessel and an uncrewed surface vehicle in the 200 kHz band is compared.

Methods
We conducted an experiment to evaluate the broadband frequency response of layers of herring by the echosounders mounted on a USV and a RV, and to document avoidance reaction, or lack of, due to an approaching platform using a bottom-moored inverted echosounder.We also modelled the frequency response of herring-like targets with different orientation angles to aid in the interpretation of the results.The experiment took place in Kvaenangen, a Fjord north of Tromsø, Norway.Here, overwintering NSS herring forms large layers suitable for these experiments (Ona et al. 2007 ).The layers are generally found at shallow depths during the night (10 −100 m) and deeper during the day (100 −300 m).Both killer whales ( Orcinus orca ) and humpback whales ( Megaptera novaeangliae ) are predating on herring (Vogel et al. 2021 ), and shallow layers and schools are known to react to vessels.This effect is particularly strong when the herring are distributed in shallow layers during the night (Ona et al. 2007 ).

Biological sampling
A standard IMR Vito pelagic sampling trawl was used for species verification and biological sampling in the observed herring layer (see, e.g.Handegard et al. 2023 , for additional imformation).Three hauls were carried out, and in each haul, 100 herring were randomly sampled from the catch, and the individual total length was measured to the < 0.5 cm.A combined total length distribution of herring was calculated where each of the three hauls was given equal weight.

Acoustic measurements
The RV and the USV used in the experiment were the 77.5m-longRV GO Sars operated by the Institute of Marine Research (IMR) in Bergen and the IMR Kayak drone (Totland and Johnsen 2022 ), respectively ( Fig. 1 ).The USV is based on a Triton expedition double kayak (Sea Kayaking, UK) equipped with an electric outboard engine, drop keel, and echosounder.The USV is programmed to follow transects between waypoints; a remote desktop connection is possible if the USV is within marine broadband radio range.An echosounder was moored at the bottom below the transects for the vessel and USV, pinging upwards to monitor the avoidance reaction of herring to the RV and USV.The transducer was positioned at a depth of 140 m.The RV and USV were equipped with Simrad EK80 echosounders (Kongsberg Discovery, Norway), whereas the echosounder mooring consisted of an autonomous Simrad EK80 WBAT (Wide Band Autonomous Transceiver).The transducers and corresponding settings used during the experiment are summarized in Table 1 for all platforms.The echosounders were calibrated using standard calibration spheres prior to the experiments according to established standards (Demer et al. 2015 ).All channels were pinging simultaneously on the RV, using suggested settings to minimize crosstalk (Khodabandeloo et al. 2021 ).
To investigate the broadband frequency response of herring distributed at various depths, the USV and RV sailed along a partially overlapping transect over herring layers over two periods, covering dusk till dawn, after the shallow herring layer had formed.The RV and USV operated close to standard survey speeds for the two platforms, ∼9 and 3 knots, respectively.The USV performed a shorter but partially overlapping transect ( Fig. 2 a).For each completed transect, we computed the broadband frequency response per platform of the herring layers measured during the transects ( Supplementary Table S1 ). Figure 2 a shows a single complete track for each of the platforms, these are indicated by full lines.The return path back to the start of the transects is indicated by hatched lines.As the experiments were performed mostly in darkness, for safety reasons, the RV was stationary during the USV transects.The RV was far enough away as to not interfere with the herring measured by the USV ( Supplementary Table S2 ), but close enough to be within radio range and have the possibility to intervene if needed.This means that the USV performed multiple tracks before each passing of the research vessel.The long track as well as the long return path back to the start of each RV transect were based on earlier experiences and WBAT observations, to ensure that the herring layer had returned to its undisturbed state and depth before the next pass of the RV.At the centre of the overlapping transects, the two platforms passed over the beam of the bottom moored echosounder, which was used to assess any avoidance reactions from the two platforms ( Fig. 2 a and  b).
After ending the avoidance and frequency response experiment with the USV and RV, and as the WBAT was operated in CW mode, the USV was used as a stationary buoy monitoring the change in broadband frequency response as the RV approached the USV.However, only one approach had a sufficient density of fish to be successful.This was performed to obtain an independent broadband measurement of herring reacting to the R V. W e also performed additional passings over the mooring with the RV alone to assess whether the patterns observed during the 24-h experiment were a general feature.

Noise and interference in the RV and USV acoustic data
Noise and interference from other sensors and equipment were observed in the acoustic data collected by the RV and USV.In the RV data, two patterns were observed in the S v ( f ) data, a significant cyclic pattern (with an approximate period of 5 kHz) and constant noise spikes.We chose to remove the spikes simply by exchanging the affected frequency with the weighted average of four adjacent frequencies (the frequency resolution in our analysis was 100 Hz), Supplementary Fig. S1 .Other potential approaches could have been a wide-band frequency subset (Echoview) or a band-pass filter (LSSS).The cyclic pattern was reduced using singular spectrum analysis (SSA), which is a nonparametric spectral estimation method T he full v ertical line indicates the position of the moored ec hosounder, the hatc hed vertical lines the start and end of a transect, the remaining parts of the echogram is from the return to the transect and is not used.Notice that this pass of the moored echosounder was not used in the analysis as the herring layer broke up temporarily.(Golyandina et al. 2021 ).Using SSA, we successfully removed the pattern by subtracting the extracted cyclic pattern from s v ( f ) , this maintained the overall shape and natural variability in the data ( Supplementary Fig. S2 ).
In the USV r ( f ) results, there is a broader peak centred around ∼198 kHz.We believe this to not represent an acoustic feature of the targets, but a that it is a more complicated combination of potentially gain (calibration), interference, and that we normalize the acoustic data [ r ( f ) ] with this region.We have chosen not to take steps to remove this, although it will likely influence the resulting statistic.In addition, TVG amplified noise due to other equipment onboard the USV prohibited us from using data deeper than the shallow herring layer.We did not perform similar processing on the integrated data from the inverted echosounder as the data was collected in narrowband mode.

Data processing and analysis
The acoustic backscatter from the moored inverted echosounder was integrated over the fish layer and is presented in s A units (nautical area scattering coefficient, m 2 nm i −2 ).The pulse repetition frequency of the echosounder was ∼1.5 Hz.To remove some of the random noise, the s A data were smoothed using a moving average with a window of 10 s, following De Robertis and Handegard (2013) .
Although the RV is equipped with additional frequencies, we focus on the 200 kHz band as this is the only overlapping frequency band between the RV and USV.The acoustic data from all mobile platforms were processed using LSSS 2.14 (Korneliussen et al. 2016 ), scripted through the LSSS API using Python 3.7.Layers were initially detected within LSSS and the layer boundaries were manually modified to ensure that the layers were herring.In addition, each herring layer recorded from the RV and the USV was manually assigned to one of two acoustic categories: 'shallow herring' or 'deep herring'.
A custom Python script was used with the LSSS API to select the data for each USV and RV transect based on recorded start and end times of each transect.S v ( f ) for 'shallow' and 'deep' herring layers was calculated for each individual transect and herring layer.The calculation of S v ( f ) in LSSS follows the procedures implemented in Simrad EK80, following Andersen et al. (2023) .The FFT window size was set to twice the length of the pulse, following recommendations by Andersen et al. (2023) , and the frequency resolution was 100 Hz.Based on the figure of merit for each transducer, extremal frequencies were trimmed at the lower and upper ends of each frequency band.For the ES200-7C transducer, this corresponds ∼10% of the frequency band at each extremity.The relative frequency response for each transect was then calculated as where s v 200 kHz is the average of s v ( f ) across the samples within the frequency band ranging from 195 to 205 kHz.To quantify any difference in r 200 ( f ) between groups, we simply integrated r 200 ( f ) from each transect across the frequency band covering both platforms, i.e.
This metric is a rather weak indicator of difference since it will not pick up any difference between the groups that averaged to zero, but in our case, it will serve as a metric of difference.To investigate the time evolution of the relative frequency response, from the experiment conducted with the USV stationary as the RV approached, r 200 ( f ) is calculated ping-by-ping and visualized as a function of frequency and time.

Acoustic modelling
A numerical modelling exercise was performed to see if we could reproduce similar frequency responses to what we observed in the in situ broadband acoustic data.The calculations of backscattering cross-section as a function of frequency and angle of incidence were performed using the Boundary Element Method (BEM ++ v. 2.0.3;Śmigaj et al. 2015 ), following the approach in Loranger et al. (2019) .As input, we used the outline of a herring swimbladder and body as provided by Gastauer ( 2023 ), 10 replicas of this artificial herring were then produced by scaling with mean herring lengths representing low and high values within the length distribution from trawl hauls in the area, 25 and 35 cm + / − 2 cm.Symmetry was assumed for body and bladder, and the outlines were converted to surface-meshed 3D models.Acoustic backscattering as a function of incidence angle and frequency of sound was calculated for the swimbladder and body separately and the results were coherently summed.The calculations were performed only at discrete tilt angles from 90 • (normal incidence of sound relative to the fish body) in steps of 5 • −25 • .For each replica, the relative target strength to a 10 kHz band ( ∼200 kHz) was calculated.The relative backscatter response was then defined as where σ bs, 200 kHz is the average of σ bs ( f ) across the samples within the frequency band ranging from 195 to 205 kHz.

Biological sampling
Three hauls were carried out using the Vito pelagic sampling trawl, and the catch of herring by haul was 2000, 60, and 600 kg, respectively.No other fish species were caught in the biological sampling.The total length distribution of herring was calculated where each of the 3 hauls was given equal weight, where the 5th percentile, median, and 95th percentile lengths were 27.0, 29.5, and 33.0 cm, respectively.

Platform avoidance monitoring with moored echosounder
The RV and the USV collected broadband data on the herring layers while performing transects passing the moored echosounder ( Table 1 ).The maximum deviation between the USV track and position of the WBAT was 3.4 m, the maximum deviation for the RV is less accurately estimated as the available logfiles from the RV contain position samples per minute.The position data are stored at a higher resolution in the raw acoustic data.However, both platforms were visible in the moored echogram data and high-resolution information on the position is not crucial for the analysis of the frequency response.
As expected, the herring layers were mainly found deep during day ( > 100 m depth) and both shallow ( ∼10 −60 m depth) and deep during night.The expected fish reaction in response to the RV and USV was confirmed using the bottom-moored echosounder.We expected shallow layers of herring to react to the RV but not the USV based on previous experiences.For the shallow layer exposed to the RV a vertical displacement was observed ( Fig. 3 a).No visible reaction was observed to the USV ( Fig. 3 b).s A averaged over all transects per platform, from 10 min before to 10 min after each pass, corroborates the visual observations; an avoidance reaction resulted in a decrease in s A when the RV passed whereas no decrease was found when the USV passed ( Fig. 3 c).We have included all passes with the RV and USV above the moored echosounder, although some passes had a very low density of fish resulting in no observable density draining due to the RV ( Fig. 3 c).This established that the response in shallow herring layers to the two platforms are different, and that there is a distinct response to the RV, which includes vertical avoidance.The vertical displacement of the herring layer in response to the RV was consistent with earlier studies in the area (Ona et al. 2007 , De Robertis andHandegard 2013 ).

Frequency response of layers
The average broadband frequency responses are estimated for each transect and platform after noise removal, as explained above ( Fig. 4 ).There is a visually distinct increase in the frequency response for the shallow herring layer when exposed to the RV, whereas for the deeper layers no such pattern is found ( Fig. 4 a).The shallow layers exposed to the USV shows a reasonable flat response across the full bandwidth for the 200 kHz transducer, and visually more similar to that of the deep layer exposed to the RV ( Fig. 4 b).
The ratio of standard deviation within each treatment groups for I ( 2 , Table 2 ) are > 4, and a non-parametric test was used to check any difference I in response to treatment.
There is a significant effect of the treatment composed of platform and layer depth (Kruskal −Wallis test, chisquared = 21.938,df = 2, and P -value < .01).A pairwise Wilcox test was used to test which of the treatments were different.Bonferroni was used to adjust the P -values for multiple comparisons.The difference between deep and shallow herring layers observed by RV was significantly different ( P = .013).In addition, the difference between deep herring observed by the RV and shallow herring observed by the USV, and the difference between shallow herring observed by the RV and shallow herring observed by the USV were both significant ( P < .001).
The moored echosounder was operated in CW mode and lacked the ability to assess the change in frequency response as the vessel approached.At the end of each of the two avoidance experiment periods, a single measurement was performed where the USV was used as a stationary buoy to monitor the frequency response as the RV approached.The echosounder data from the stationary USV showed a similar pattern as the moored echosounder ( Fig. 5 a).As the herring dives, the S v ( f , t ) decreases across all frequencies ( Fig. 5 b), and the r 200 ( f , t ) shows a similar pattern as seen in the RV data in the upper part of the 200 kHz frequency band ( Fig. 5 c).This further supports the evidence for avoidancedependent broadband frequency response and illustrates the potential to use r ( f , t ) to monitor platform-induced avoidance or vertical swimming behaviour in fish in general.Unfortunately, only one experiment was successful, as the herring density was too low in the second experiment to extract a meaningful r 200 ( f , t ) .duce a flat frequency response at normal incidence of sound and to a degree an increase in σ bs > 200 kHz, but the onset of the increase is found at higher frequencies than in the measurements.The increase begins to become prominent at 50 • tilt.Figure 4 c shows the averaged relative frequency response for 50 • −90 • tilt (normal incidence of sound).

Discussion
We have shown that the broadband frequency response in NSS herring is different when comparing the observations from a research vessel and an USV, and that the difference can be attributed to changes in vessel-induced fish behaviour.Based on Note that pairs in deep and shallow herring are calculated from the same transect.The reason for shallow herring to be less is caused by lack of fish in the upper layer on one passing.The summary (mean, SD , median, and inter quartile range) for the integrated frequency response I ( 2) is also provided.previous studies, the RV induced an expected behaviour response in shallow layers of herring, whereas the fish did not react to the USV.Note that no observations on the deeper layers were available from the USV.It is worth emphasizing that Norwegian spring spawning herring is overwintering in the area where our experiments took place.They are partic-ularly reactive at nighttime when forming shallow layers.As discussed in, e.g.papers covering vessel avoidance (De Robertis and Handegard 2013 ), it is not possible to generalize this behaviour to other species, geographical areas, or times of year.Fish reacting to an approaching research vessel may bias acoustic surveys, and traditionally, we have used experiments to assess any effects (Ona et al. 2007, De Robertis and Handegard 2013, Evans et al. 2023 ).These approaches are not the most cost-effective approaches as they require dedicated vessel time and/or equipment, or do not directly assess potential bias in the data collected for assessment.Another approach is to use sonar (Soria et al. 1996 ), but that typically covers different parts of the water column and not directly applicable to echosounder surveys.If the frequency response from avoiding/not avoiding herring can be established, then this response can be used to assess any potential effects as the data is being collected.This is typically useful when there are variations in the response with potential annual bias.
RV GO Sars performed additional transects across the herring layers, also in adjacent fjords, without the USV and mooring, yielding comparable frequency responses as found during the experiment described in this paper.This confirms that it was a persistent feature across the fjords during our experiments.The frequency response obtained by the RV of deep and shallow herring in the lower frequency bands, 34 −160 kHz, did not display obvious differences attributable to fish avoidance ( Supplementary Fig. S3 ).Observations of a comparable rise in the relative frequency response with the USV as fish avoid the approaching RV ( Fig. 5 ) documents that this effect is observable on other platforms than the R V. W e have recently also experienced similar differences between GO Sars and other USVs (unpublished data).
The broadband frequency response of herring distributed in shallow waters observed with the USV are visually similar to that observed from the RV for herring distributed in deep waters ( Fig. 4 ).We do note that there are some differences between the frequency response of the deeply distributed herring and the surface layer, where the frequency response for the undisturbed surface layers has a slightly negative slope at high frequencies.The cause is not obvious; if anything, the slope could be expected to be flatter as the fish bladders are likely less compressed, and further from fundamental resonance, in the shallow relative to the deep layer.Earlier studies have shown that herring distributed at depth compensate for the compressed swimbladder, and hence negative buoyancy, by altering the swimming angle (Huse and Ona 1996 ).Herring at depth has been found to have a positive swimming angle or a positive swim-negative glide pattern, depending on time of day.This could lead to similar effects on the frequency response as caused by the avoidance of disturbed shallow layers.However, we also note that the difference between avoiding and undisturbed herring in the shallow surface lay-ers are much larger, indicating the avoidance effect is a much stronger behavioural response than the potential swim-glide patterns.Differences in frequency response may also be a direct function of depth, an additional effect that needs attention when developing automated methods on acoustic data.In addition, there are still likely residual artefacts in the data, such as the peak < 200 kHz in the USV data ( Fig. 4 a), which introduces some uncertainty in the interpretation of the results.Additional datasets on deep and shallow herring, which are less affected by noise would be required to conclude whether there is a significant difference between the undisturbed shallow and deep cases.
As described in the Methods section, we investigated and took steps to reduce noise and interference, which could introduce biases in r ( f ) and hence the interpretation of the results.Based on this scrutiny, as well as the independent observations by the USV as a stationary buoy, biases are unlikely to cause the high-frequency increase in r ( f ) for the shallow layers observed by the RV.However, we cannot rule out that differences or similarities between shallow USV and deep RV r ( f ) may be attributable to artifacts.This would require new data collected with less interference from onboard equipment.
The model results indicate a flat frequency response at normal incidence of sound and an increase > 200 kHz, but the onset of the increase is found at higher frequencies than in the measurements; in addition, the variation in r 200 ( f ) and the magnitude at higher frequencies are larger than observed.Nevertheless, this provides further support of the hypothesis that behaviour alterations (avoidance) and consequently changes in tilt and swimming angle contribute to what we observe in the measurements.Potential weaknesses in our current modelling approach included not considering depth effects on the physostome swimbladder, whether the volumes and shapes of the swimbladders we have used are appropriate for the depths they occupy, disregarding other internal structures such as bone, and coherently summing the contributions from bladder and body.With improvements in the model parameters, we could likely improve our capability of predicting the frequency-dependent backscattering of physostome fish avoiding platforms.An improved modelling approach combined with controlled measurements would help in understanding the mechanisms behind the observed frequency response, and potentially aid in quantifying behaviour from the broadband frequency response.Although we have not taken steps to test the consequences of differences in frequency response between platforms on classifiers, the differences we observe will likely affect and potentially bias classifiers.

Conclusions
The USV Kayak does not affect shallow herring layers; the RV GO Sars does, based on acoustic observations by an inverted echosounder.The r 200 ( f ) of shallow herring observed by the RV is significantly different than the response observed by the USV.The mechanisms behind the effect on the broadband frequency response is still unclear and requires further investigation.Creating acoustic classification methods for physostome fish care should be taken as the frequency response may differ between observation platforms.Monitoring r ( f ) of layers of physostome fish can potentially be used to assess whether the layers are responding to the vessel ( Fig. 4 ) and as a general tool for monitoring fish behaviour.

Figure 2 .
Figure 2. Experimental setup.(a) Example of repeated transects with RV (long transect) and USV (short transect) to measure frequency response of shallow and deep herring over the autonomous moored echosounder (circle).The hatched lines indicate return to the transect after finishing a transect.The direction of the transects was from north to w ards south.(b) Example echogram collected by the R V sho wing shallo w (at ∼10-60 m depth) and deep herring la y ers.T he full v ertical line indicates the position of the moored ec hosounder, the hatc hed vertical lines the start and end of a transect, the remaining parts of the echogram is from the return to the transect and is not used.Notice that this pass of the moored echosounder was not used in the analysis as the herring layer broke up temporarily.
Again focusing on the joint USV and RV 200 kHz frequency band, we calculate the average relative frequency response [ r 200 ( f ) ], relative to σ bs at 200 kHz.The model can repro-Downloaded from https://academic.oup.com/icesjms/advance-article/doi/10.1093/icesjms/fsae048/7660037 by guest on 11 May 2024

Figure 3 .
Figure 3. Observations made by the moored echosounder.Distribution of the herring layer observed by the moored echosounder when passed directly abo v e b y the R V (a) and the USV (b).T he time when the R V and USV passed the moored echosounder is indicated with v ertical hatched lines.T he w ak e of the RV is also visible in (a).(c) Moored echosounder sA values for each individual pass, 10 min before and after each pass are shown with light lines.T he a v erage sA per platf orm is indicated b y the full lines.T here is a clear visual response to the R V and a lack of a v oidance reaction to the USV.

Figure 4 .
Figure 4. (a) The relative frequency response r 200 (f) for each transect from the RV and USV transducers that o v erlaps in frequency.T he curv es sho w deep (GOS.DeepHerring) and shallow (GOS.ShallowHerring) herring la y ers, observ ed from the R V, and shallo w la y ers from the USV (Ka y ak.Shallo wHerring).(b) The integrated log transformed r 200 (f).(c) Averaged modelled relative frequency response for 50 • −90 • tilt (normal incidence of sound).

Figure 5 .
Figure 5. (a) Avoidance reaction by a herring layer as the RV approaches the USV, as observed by the USV 200 kHz echosounder while the USV is held stationary.The closest point of approach is between 23:00 and 23:01.From the echogram, a clear reaction is observed due to the approaching RV.(b) Time evolution of the 200 kHz broadband Sv frequency response as the RV approaches the USV and the herring dives.The figure shows multiple broadband spectra for the herring layer combined to produce a broadband echosounder 'spectrogram', showing Sv as a function of time and frequency.(c) Time e v olution of the relative frequency response (re 200 kHz) showing an increase in the higher part of the 200 kHz frequency band, as we also observed from the research vessels' echosounders ( Fig. 4 ), as the herring dives.

Table 1 .
The operating frequencies and power settings for the RV, USV, and moored echosounder.
PlatformTransciever Transducer Mode Frequency range (kHz) Pulse duration (ms) Po w er (W)