Uncrewed surface vehicle (USV) survey of walleye pollock, Gadus chalcogrammus, in response to the cancellation of ship-based surveys

Abstract In 2020, the developing COVID-19 pandemic disrupted fisheries surveys to an unprecedented extent. Many surveys were cancelled, including those for walleye pollock (Gadus chalcogrammus) in the eastern Bering Sea (EBS), the largest fishery in the United States. To partially mitigate the loss of survey information, we deployed three uncrewed surface vehicles (USVs) equipped with echosounders to extend the ship-based acoustic-trawl time series of pollock abundance. Trawling was not possible from USVs, so an empirical relationship between pollock backscatter and biomass established from previous surveys was developed to convert USV backscatter observations into pollock abundance. The EBS is well suited for this approach since pollock dominate midwater fishes in the survey area. Acoustic data from the USVs were combined with historical surveys to provide a consistent fishery-independent index in 2020. This application demonstrates the unique capabilities of USVs and how they could be rapidly deployed to collect information on pollock abundance and distribution when a ship-based survey was not feasible. We note the limitations of this approach (e.g. higher uncertainty relative to previous ship-based surveys), but found the USV survey to be useful in informing the stock assessment in a situation where ship-based surveys were not possible.


Introduction
The COVID-19 pandemic resulted in the cancellation of many fisheries surveys worldwide in 2020. This posed a challenge for fisheries management, which relies on timely and consistent abundance estimates of fish stocks to characterize the state of marine ecosystems to support management decisions (ICES, 2020). This was the case for walleye pollock (Gadus chalcogrammus) in the eastern Bering Sea (EBS), which support the largest single-species fishery in the United States with recent landings of ∼1.3 million tons and a value of ∼1.4 billion dollars (Ianelli et al., 2020). The ship-based acoustictrawl (AT) and bottom trawl surveys used in the stock assessment (Ianelli et al., 2020) were delayed and subsequently cancelled due to the risk of the virus to the survey crews and the remote communities where crew exchanges and resupply activities occur. In response, we applied recent advancements in uncrewed surface vehicles (USVs) instrumented with calibrated echosounders (e.g. Swart et al., 2016;De Robertis et al., 2019) to conduct a USV-based acoustic survey. The goal was to mitigate the loss of information from pollock midwater abundance surveys used to support management of this important fishery.
In recent years, ocean-going autonomous vehicles and sensors suitable for measuring the abundance of marine organisms have proliferated (Fernandes et al., 2003;Verfuss et al., 2019;  A. De Robertis et al. Reiss et al., 2021). USVs using wind and wave power for propulsion are promising for large-scale fisheries surveys (Greene et al., 2014;Swart et al., 2016) due to their long endurance. Initial short-term deployments (Greene et al., 2014) have been extended to >100 days (De Robertis et al., 2019;Chu et al., 2019;Levine et al., 2021), and USVs capable of conducting acoustic surveys are now commercially available (Verfuss et al., 2019). Low-power echosounders suitable for acoustic measurement of fish abundance from USVs have also been developed (Benoit-Bird et al., 2018). These USV systems have matured to the point that, in many cases, the focus of work is transitioning from the development of vehicles and sensors to how these new tools can best be applied to understand the abundance, distribution, and ecology of marine organisms.
The primary limitation of USVs for acoustic surveys of fish abundance is their inability to perform concurrent direct biological sampling to identify acoustic targets (Fernandes et al., 2003). Acoustic methods are currently insufficient to reliably distinguish among fish species and size classes in most operational situations (McClatchie et al., 2000;Bassett et al., 2017;Korneliussen, 2018). As a result, targeted trawl samples are generally used to identify species and size/age composition of acoustic scatterers during ship-based surveys (McClatchie et al., 2000, Simmonds andMacLennan, 2005). However, in some instances, diversity may be sufficiently low that a single species or group dominates acoustic backscatter and species identification may be less of a concern (Geoffroy et al., 2011;Ressler et al., 2015;Reiss et al., 2021). In the case of the EBS, age 1 + pollock dominate acoustic scattering in deeper parts of the water column to the point that this backscatter can be attributed entirely to pollock (De Robertis et al., 2010;Honkalehto et al., 2011). As a result, acoustic-only measurements in this environment can be translated into a useful index of total pollock abundance (note that this differs from indices of abundance by size/age produced by traditional surveys). For example, an index of total abundance based on backscatter measurements (i.e. no trawl data) collected opportunistically during the bottom trawl survey has been used as an input to the EBS pollock stock assessment since 2011 (Honkalehto et al., 2011). It thus appeared feasible to generate a useful index of pollock abundance from USV acoustic measurements without concurrent trawl sampling.
Here, we describe an effort to apply these recent technological developments to mitigate the loss of information resulting from survey cancellations caused by the COVID-19 pandemic in summer 2020. The goal was to conduct a USV survey to produce an abundance index consistent with that from a conventional ship-based AT survey in time for incorporation into the annual stock assessment. We used three chartered saildrone USVs to estimate the abundance of walleye pollock on the EBS shelf and extend an AT survey time series ordinarily conducted with a research vessel (McCarthy et al., 2020). The primary challenges faced in this effort were (1) the lack of biological information from trawl sampling, (2) the short timeline to provide information for the stock assessment, and (3) generating results comparable to those from conventional ship-based AT surveys with USVs. We describe the rapid and successful adoption of USVs to mitigate the loss of information from a cancelled shipbased AT survey, and highlight considerations relevant for future application of USVs in fisheries surveys.

Approach
The 2020 AT survey of pollock in the EBS was cancelled due to safety concerns associated with the COVID-19 pandemic. Instead, three chartered saildrone USVs were deployed from Alameda, CA, to the Bering Sea to estimate pollock abundance and distribution. The USVs surveyed the same area (3.5 × 10 5 km 2 ) as recent AT surveys, but transects were spaced farther apart due to time constraints. The USVs measured acoustic backscatter, but population biomass (kg) is used in the stock assessment model. Thus, the USV backscatter measurements were converted to biomass units based on an empirical relationship between pollock backscatter and biomass derived from previous surveys. Finally, the additional uncertainty introduced by the increased transect spacing and the backscatter-to-biomass conversion was investigated via simple simulations.

USV survey
Three saildrone USVs equipped with echosounders were used to conduct the survey. Saildrones are 7 m long wind-propelled and solar-powered USVs equipped with oceanographic and meteorological instruments (Mordy et al., 2017). The vehicles autonomously follow pre-planned routes and can be monitored and re-tasked via a satellite link.
The USVs followed a curtailed survey plan designed in case an abbreviated ship-based survey had been possible. The survey consisted of 14 transects spaced 74 km apart with a total length of 4727 km (Figure 1). This represents half the sampling density of previous surveys (transects 37 km apart). The transects extended from water depths of ∼60-80 m to the shelf break (>1000 m). The average depth was 127 m and 90% of the area was between 65 and 287 m deep. The survey area was divided into three sectors, each covered by a USV proceeding from southeast to northwest. As in previous surveys, the survey was conducted during daytime; the USVs held position between sunset and sunrise. The USVs paused at wind speeds > 12 m s -1 as wind-mixed air bubbles attenuate the echosounder transmissions at high wind speeds (De Robertis et al., 2019).
Due to travel restrictions, the USVs could not be deployed near the survey area. Instead, they departed on 15-22 May from Alameda, California and sailed 3700 km across the North Pacific to the Bering Sea ( Figure 1). The USVs began the survey on 4 July and finished on 20 August, within the date range of previous acoustic surveys in the EBS ( Figure S1). After completing the survey, the USVs sailed back to Alameda. The USVs were recovered on 4 October and raw acoustic data were sent by overnight courier for further processing.
The USVs were equipped with a Simrad WBT mini split-beam echosounder equipped with an ES38-18/200-18CR transducer (18 • split-beam 38 kHz and 18 • single beam 200 kHz) gimbal mounted on the keel at a depth of 1.9 m (De Robertis et al., 2019;Levine et al., 2021). The echosounders continuously transmitted 0.512 ms narrowband signals every 3 sec. Although backscatter data were collected to 1000 m in previous surveys, only a small amount of pollock backscatter was observed >300 m (mean 0.32%, range 0.08-0.67%). We thus chose to record USV data to 300 m to simplify data collection as this required fewer changes to instrument settings during the survey. Raw echosounder data were written to flash memory media, and data summaries were transmitted over the satellite link (De Robertis et al., 2019).
The echosounders were calibrated before and after the survey in San Francisco Bay on 13-14 May and 28 October 2020 following Renfree et al. (2019). These calibrations were conducted at warmer temperatures (16.7/18.3 • C at the transducer) than encountered during the survey at the transducer depth (mean 10.3 • C, 95% of observations 8.3-12.8 • C), which increases uncertainty due to the potential for temperature-dependent gains (Demer and Renfree, 2008). The mean gain (averaged in the linear domain) from the two calibrations was used in data processing. The pre-and post-survey calibration results were consistent, with gain differing by < 0.1 dB at 38 kHz and < 0.2 dB at 200 kHz for all USVs.

Processing of acoustic data
The acoustic data were processed using Echoview v11.0 software. Pings exhibiting evidence of attenuation of the transmitted signal by bubbles entrained into surface waters at elevated sea states (i.e. a locally weak bottom echo or variable transmit pulse) were removed following the criteria described in De Robertis et al. (2019). The data were processed by four analysts who, as in previous surveys, visually inspected 38-kHz echograms and subjectively attributed backscatter to pollock based on characteristics such as depth distribution (often near-bottom), appearance (consistent with individual fish or schools), frequency response (similar at 38 and 200 kHz), strength of the returned echo (high), and geographic location (scarce in shallow areas). In most cases, acoustic backscatter consistent with fish aggregations in the lower one-third to twothirds of the water column is dominated by age 1 + walleye pollock targeted by the survey, and analysts primarily demarcate the depth below which backscatter is attributed to adult pollock (De Robertis et al., 2010;Honkalehto et al., 2011).
In some areas, pollock co-occur with diffuse backscatter from other scatterers making discrimination of pollock and other backscatter more challenging. As in previous surveys (McCarthy et al., 2020), the analysts had the option of applying a two-part filter to limit the influence of these scatterers prior to assigning backscatter as consistent with pollock. This filter applied a -60 dB re 1 m -1 threshold to limit the influence of weak targets unlikely to be pollock and excluded backscatter with a frequency response outside of 2 standard deviations of that expected for pollock (i.e. S v,38khz -S v,200 kHz > 7.9 or < -0.9 dB was excluded; De Robertis et al., 2010). This was applied to 8.5% of the survey and 3.5% of total pollock backscatter was processed in this fashion. Acoustic backscatter attributed to pollock (hereafter pollock backscatter) from consecutive 50-ping intervals was echo-integrated using an integration threshold of -70 dB re 1 m -1 into the nautical area scattering coefficient (s A, m 2 nmi -2 ) in the two layers used in previous surveys: >3 m (16 m below the surface to 3 m above the bottom detection), and <3 m (3 to 0.5 m above the bottom detection).
The course of saildrones depends on wind speed and direction as they must tack back and forth to sail upwind, and they drift when becalmed. The distance traveled on a given pre-planned transect thus depends on the wind (e.g. a longer distance will be traveled upwind due to the indirect path), and each along-track interval thus represents a variable distance along the intended transect. We accounted for this potential bias by averaging the processed USV observations into 0.5 nmi (926 m) along-track elementary distance sampling units (EDSUs) defined along the pre-planned intended transects so that each EDSU represented a consistent distance along the intended transect. More 50-ping intervals were associated with each EDSU at low speeds and when traveling upwind. On average, 7.6 intervals were associated with each EDSU (90% of EDSUs in the range of 3-17 intervals). The average number of pings per EDSU for the saildrones (∼380) is comparable to the number of pings transmitted by a ship transiting at 6 m s -1 (300).

Adjustment for USV beam width
The USVs were equipped with wider beam transducers (17.0-17.9 • ) than the 7.0 • transducers used in previous surveys. A wider beam leads to a larger near-bottom acoustic dead zone (ADZ) where backscatter from fish cannot be discriminated from the seafloor. The USVs are thus expected to detect less backscatter near the seafloor than the research vessel. The magnitude of this bias depends on the range to the seafloor, the degree to which fish aggregate near the seafloor, and the transducer beam width (Ona and Mitson, 1996). The unobserved backscatter in the ADZ was estimated by computing the mean height of the ADZ and extrapolating the backscatter observed immediately above the ADZ (i.e. the layer extending from 0.5 to 2m above the bottom echo) into the unobserved volume using the method of Ona and Mitson (1996). The survey time series does not include such a correction for the ADZ. Thus, USV observations were corrected to the ADZ expected for the 7.0 • transducer used in previous surveys to maintain consistency with earlier results. This was accomplished by adding the difference between the backscatter estimated in the ADZ for the USV and the research vessel (RV) to the USV observation in the 0.5 to 3 m above bottom layer: (1) Survey-wide, this correction increased total pollock backscatter by 3.2%.

Backscatter to biomass conversion
Pollock biomass is the primary quantity of interest for stock assessment. Trawling was not possible during the USV survey, and measurements of pollock backscatter needed to be converted to biomass units without trawl sampling. We thus examined 16 previous AT surveys of pollock in the EBS conducted between 1994 and 2018 (McCarthy et al., 2020) to determine whether a consistent relationship to convert between pollock backscatter and biomass could be derived.
Pollock dominated the weight of fish caught in pelagic trawls during these surveys (mean = 98.1%, range 95.2-99.8%). Note that large medusae, mostly Chrysaora melanaster, which accounted for 7.4% (range 1.0-16.4%) of total catch weight were excluded from the calculations as they are negligible contributors to backscatter relative to fish with swimbladders (De Robertis and Taylor, 2014). Pollock also dominate catches in the summer commercial fishery which is conducted largely in the survey area (Ianelli et al., 2020): pollock accounted for an average of 98.8% of total catch by weight from 2000-2020 (range 97.6-99.4%). In summer 2020, when the USV survey was conducted, the proportion of pollock in the fishery catch was 97.8%. Together, these catch records suggest that acoustic scattering from fish aggregations in this region of the Bering Sea is dominated by pollock.
In the AT surveys, pelagic observations (>3 m) are treated differently than demersal observations (<3 m). Targeted midwater trawl samples are used to convert pelagic backscatter from 16 m below the surface to 3 m above the seafloor to fish abundance (Mc-Carthy et al., 2020). Demersal backscatter, where fishes are more diverse and pollock tend to be larger, is converted to biomass based on species and size composition from a concurrent bottom trawl survey and estimates of species-specific contributions to acoustic backscatter (Lauffenburger et al., 2017). Pollock lengths from trawl catches are used to convert backscatter to fish abundance using a target strength to length relationship (Traynor, 1996), and biomass is determined by multiplying abundance by weight-length relationships determined from the trawl catches.
For each survey, indices of the total acoustic backscattering cross section attributed to pollock [A T = i s A,i 4π E i , (m 2 ), hereafter referred to as total backscatter] and the total pollock biomass [B T = i B i E i , (kg)] were computed from observations of s A (m 2 nmi -2 ), or biomass density (kg nmi -2 ), at each 0.5 nmi (926 m) EDSU, i, and the area represented by each EDSU E i (20 nmi 2 ). As previous surveys exhibited a consistent relationship between total pollock backscatter, A T , and biomass, B T , linear regressions with an intercept of zero (see results for details) were used to estimate C (kg m -2 ), a unit conversion used to convert the 2020 USV backscatter observations to pollock biomass (i.e. B T = C × A T ). As in previous surveys, the midwater (>3 m), and demersal (<3 m) backscatter observations were converted to biomass units using separate relationships.
To examine the potential prediction errors associated with this conversion, we conducted a leave-one-out cross-validation exercise in which the biomass in each survey year was predicted based on the backscatter to biomass relationship observed in the other years. In each year, the survey biomass estimate (i.e. backscatter converted using trawl samples) was compared to that estimated using backscatter to biomass conversions established with observations from the other years as described above.

Uncertainty estimates
Uncertainty in the EBS pollock survey is quantified using a 1D geostatistical method based on transect biomass sums (Williamson and Traynor, 1996). The resulting "relative estimation error" accounts for uncertainty due to unobserved spatial structure and measurement error on the abundance estimate. If this uncertainty can be assumed to be normally distributed, it provides an estimate of the coefficient of variation (i.e. CV = σ /K) where K is the estimated abundance and σ is the square root of the estimation variance associated with the degree of spatial sampling. This assumption is warranted for the EBS AT survey as two times the 1D CV results in an approximate estimate of the 95% confidence interval of precision for this survey (Walline, 2007). We followed this approach to produce an uncertainty estimate comparable to those from previous surveys. Given that fewer transects were sampled in 2020 than in previous surveys, and that backscatter was converted to biomass based on the historical survey results rather than directed trawl samples, one would expect the resulting uncertainties to be larger in the 2020 USV survey.
The 1D method was applied to backscatter in the >3 m and <3 m layers independently. The results were combined assuming independent variances [as the estimated variances in previous years from the 1D method computed for >3 m and <3 m were not significantly correlated (r = 0.21, p > 0.40)]: where σ 2 is the estimated variance and A T is total pollock backscatter.
We characterized the potential effects of using 74 km rather than 37 km transect spacing in the 2020 survey by comparing backscatter estimates from previous surveys calculated using half the transects (i.e. even or odd numbered transects) with estimates using all transects (37 km spacing). The absolute % difference between the total backscatter at all and half transect spacing was used to characterize the discrepancies. For each survey, we computed the 1D CV for backscatter measurements based on all transects (i.e. 37 km transect spacing as in previous surveys), and compared this to the result of using only the odd-or even-numbered transects (i.e. 74 km spacing as in the 2020 survey). The 2020 USV survey lacked trawl sampling and thus incorporated a backscatter-to-biomass conversion based on the average relationships observed in previous surveys. We accounted for the additional uncertainty introduced by this conversion by propagating the estimated uncertainty in the conversion using a Monte-Carlo simulation. We first drew a random sample j of total backscatter based on the estimated survey backscatter A T and its uncertainty (i.e. A T j is drawn from a normal distribution with mean of A T and a SD corresponding to A T multiplied by the 1D CV). This step accounts for uncertainty associated with spatial sampling. We then added additional uncertainty to account for the backscatterto-biomass conversion. Specifically we sampled C j from a normal distribution with a mean corresponding to C and a SD corresponding to the standard error of the prediction interval from the regression described in the results section. This step accounts for the uncertainty in the biomass that we might have observed in 2020, given the backscatter to biomass relationships from previous years. Biomass is computed as This process was repeated for each of 10 5 replicates j. The CV of the USV survey biomass was estimated from the mean and standard deviation of all replicates B T j . This approach accounts for the reduced sampling in 2020 (as 1D CV will be higher) and the additional uncertainty introduced by the backscatter to biomass conversion.

USV survey
The USVs completed the survey without issue. Wind speeds during the survey were modest, averaging 5.4 ± 2.5 m s -1 (mean ± SD) with 95% of wind speeds < 9.7 m s -1 . The USVs travelled at an average speed of 0.9 ± 0.5 m s -1 , achieving a maximum speed of 3.0 m s -1 . Acoustic data were collected throughout the deployment. The geographic distribution of acoustic backscatter was similar to that in previous surveys ( Figure S2). The vertical distribution was also consistent with previous surveys (De Robertis et al., 2010;Mc-Carthy et al, 2020). In general, there was a layer of shallow backscatter that was not attributed to pollock, and a near-seafloor layer of backscatter that was attributed to age 1 + pollock (Figure 2a). In some cases, scattering was attributed almost entirely to age 1 + pollock (Figure 2b), and in others, scattering from other sources occupied most of the water column (Figure 2c). The strong backscatter from fish schools near the surface (e.g. <40 m in Figure 2a  The lower panels show the relationship between total pollock backscatter (A T ) and biomass (B T ) computed based on trawl sampling (b) > m above bottom and (c) . to m above bottom layers. Each point represents a different year. The black line represents a linear regression, with gray shading representing the % confidence interval for the expected relationship, and the gray dotted lines represent the % prediction interval of the regression (i.e. confidence intervals for a future observation).
trawl-mounted cameras indicates that age-1 + pollock are scarce in these two scattering layers Williams, 2018, Mc-Carthy et al., 2020). Survey-wide, 45.0% of the backscatter was attributed to pollock, which is similar to previous surveys (48.6 ± 16.5%, range 23.0-85.3%). As in previous surveys, pollock were more abundant near the seafloor, with few pollock detected near the surface. The majority of backscatter >70 m was attributed to pollock (Figure 2c).

Backscatter to biomass conversion
Pollock backscatter in previous surveys exhibited a similar trend as pollock biomass (Figure 3a). This suggests that it may be possible to estimate survey-wide pollock biomass reasonably well without trawl information. Linear regression of survey-wide pollock backscatter (A T ) and pollock biomass (B T ) indicates that backscatter is well correlated with biomass in both the >3 m layer (Figure 3b, B T = 1053.7 kg/m 2 ·A T , r 2 = 0.99, p < 0.001), and the <3 m layer (Figure 3c, B T = 1043.3 kg/m 2 ·A T , r 2 = 0.99, p < 0.001).
Cross-validation revealed that in previous surveys, pollock biomass could be estimated reasonably well with this conversion approach without trawl information in a single year (i.e. Figure 3b and c). Overall, the survey biomass estimated using the conversion exhibited a comparable trend as the conventional survey biomass (Figure 4, r 2 = 0.97). The average absolute value of the discrepancy between the survey and the converted estimate was 5.6%, with a range of 0.1 to 12.3%.
Acoustic backscatter from only odd or even transects from previous surveys (74 km spacing) produced similar results as all transects (37 km spacing, Figure 5a), suggesting that the 74 km spacing used in the USV survey was sufficient to capture the primary trends in the abundance time series. The average absolute discrepancy in survey-wide backscatter at 74 km and 37 km spacing was 3.9 ± 2.5% (mean ± SD). The average 1D CV from 74 km spacing was 0.097 ± 0.039, which is 2.0 times larger than the average CV (0.049 ± 0.020) in previous years when the survey was conducted at 37 km spacing (Figure 5b

USV abundance estimates
Total pollock backscatter in the survey area (A T ) was 3.44 × 10 6 m 2 , 45.0% higher than in the last survey in 2018 (Figure 3a). The spatial distribution of pollock backscatter was consistent with recent surveys, with pollock most abundant in the northwest portion of the survey area ( Figure S2). The biomass estimate for 2020 was 3.62 × 10 9 kg of pollock, which represents an increase of 44.7% relative to the estimate of 2.49 ×10 9 kg in the last survey in 2018 ( Figure 6). The CV based on the 1D geostatistical estimate for backscatter measurements in 2020 was 0.068, which is similar to previous surveys reduced to 74 km spacing (Figure 5b). The CV increased to 0.096 when the additional uncertainty of the backscatter to biomass conversion was incorporated (Figure 7). Thus, for the 2020 survey, the uncertainty attributed to spatial sampling is about twice that of the backscatter to biomass conversion. The overall CV is 2.3 times the average CV of 1994-2018 time series (0.042, Figure 6). The proportion of backscatter (15.3%) in the lower 3 m was in the range observed in previous years (11.7-30.2%, Figure S3), which suggests that after correction, the difference in beam width between the USV and the research vessels used for previous surveys did not introduce major bias.

USV pollock survey
This work used USVs instrumented with echosounders (Swart et al., 2016;De Robertis et al., 2019) to continue a survey time series after the unexpected cancellation of an AT survey. Multiple USVs were used to perform the survey within the time window of previous ship-based surveys to minimize the impact of potential stock migration on survey results. Previous experience working with these USVs (De Robertis et al., 2019;Meinig et al., 2019) allowed for the rapid development and execution of the survey as it became clear that the vessel-based surveys could be cancelled. Long-term survey time series are a fundamental tool in fisheries management, allowing for fish stock abundance and distribution to be assessed over time (Hilborn and Walters, 2003). Therefore, it is essential that changes to survey methodology be considered in relation to the existing time series, so that direct comparisons to previous surveys remain possible. Where possible, the ship-based AT survey methodology (McCarthy et al., 2020) was followed to produce a comparable USV estimate. The lack of concurrent biological sampling remains the primary limitation to using acoustic data from USVs to inform fisheries management (Fernandes et al., 2003). In AT surveys, targeted trawl samples are used to allocate backscatter to species and size classes (Simmonds and MacLennan, 2005), which is not feasible from USVs. In some circumstances (e.g. low-diversity high latitude ecosystems) acoustic backscatter can be attributed to species or important taxonomic groups without additional biological sampling (Geoffroy et al., 2011;Ressler et al., 2015;Reiss et al., 2021). This is the case in the EBS survey area, where pollock account for almost all the backscatter from semi-demersal fishes (De Robertis et al., 2010), and an acoustic-only index is used in the stock assessment (Honkalehto et al., 2011;Ianelli et al., 2020).
The USV survey should not be considered a complete replacement for the AT survey as it does not provide all the information produced by the ship-based survey. For example, estimates of size/age composition and biological measurements from fish specimens, which are important components of this survey (Ianelli et al., 2020) were not possible. The 2020 biomass estimate is also more uncertain than those from previous surveys due to reduced survey coverage and the backscatter to biomass conversion. Most of the increased uncertainty, in this case, is associated with the increased transect spacing rather than the biomass conversion. In addition, the backscatter to biomass conversion approach relies on the assumption that the species and size composition of fishes during the USV survey falls within the range observed in previous surveys. Given these limitations, we plan to conduct future AT surveys of EBS pollock with crewed vessels. However, the USV survey met the intended goal of providing quantitative information on the abundance and distribution of the pollock stock when ship-based surveys were not possible.

Application to fisheries management
The abundance at age index from the AT survey is critical for providing near-term advice in the stock assessment because historical age compositions indicate that it uniquely provides reliable estimates of "pre-recruit" fish. These younger age groups tend to be more pelagic and more available to the AT survey than bottom trawl surveys (Ianelli et al., 2020). This means that the AT survey provides the first sign of pollock abundance between the ages of 2-4 (full recruitment to the fishery generally is around age 5). Additionally, the AT survey provides a fishery-independent snapshot of pollock distribution throughout the region. This has become increasingly important in tracking the impact of warming and more variable environmental conditions within the Bering Sea (Stevenson and Lauth, 2019;Eisner et al., 2020).
In 2020, the USV survey was the only source of fisheriesindependent data available for the EBS pollock stock to the North Pacific Management Council (NPFMC) as the scheduled AT and bottom trawl surveys were cancelled (Ianelli et al., 2020). To extend the existing AT time series with the USV data, it was important to convert the USV-measured acoustic backscatter to total biomass and quantify the additional uncertainty introduced by the use of USVs as described above. Adding the USV data to the assessment model provided assurance that the stock status was stable and suggested a slight increasing trend compared to the previous survey and other model scenarios (Ianelli et al., 2020). The USV data were broadly consistent with other data components fit within the assessment model. Furthermore, the pollock spatial pattern depicted by the USV survey was consistent with the patterns observed in the summer fishery providing additional confidence in the USV results. The assessment model scenario incorporating the USV data was selected by the NPFMC as the basis for management advice.

Backscatter to biomass conversion
In this study, we derived simple empirical relationships to convert USV backscatter measurements into total pollock biomass. This conversion was possible because pollock account for most of the observed backscatter from semi-demersal fishes, particularly in the >3 m above the bottom layer where 85% of pollock backscatter was observed in this survey. Given that age-1 + pollock dominate acoustic scattering from semi-demersal fishes (De Robertis et al., 2010;Honkalehto et al., 2011), much of the variability in the backscatter to biomass conversion is attributable to changes in pollock size distribution. The conversion also exhibits low variability in the <3 m layer where fishes are more diverse, indicating that pollock also account for a high (∼80%) and consistent proportion of backscatter across years in this layer (Lauffenburger et al., 2017).
The approach used in this study is equivalent to applying average mass-specific target strengths to convert backscatter to biomass. Backscatter from individual pollock scales as length 2 (Traynor, 1996) and mass scales approximately as length 3 (e.g. McCarthy et al., 2020). Thus, mass-specific scattering is size-dependent and scales approximately as length -1 . The backscatter-to-biomass conversion for pollock can be thought of as applying an average species and size composition in the absence of trawl samples. When expressed as a mass specific target strength (TS = 10·log 10 (C -1 )) the conversions in this study (−30.2 dB m 2 kg -1 for both <3 m and >3 m) are very similar to previous estimates derived from target strength measurements and used in early EBS AT surveys (-30 dB re 1 m 2 kg -1 , Karp and Walters, 1994). The conversion approach used here is potentially applicable to other situations where backscatter can be attributed to species or taxonomic groups without trawl sampling and the average mass-specific target strength is relatively consistent. Existing survey data can be used to estimate the additional uncertainty introduced if a backscatter-to-biomass conversion based on previous surveys is applied (Figure 3).
The consistent relationships between observed backscatter and survey estimates of biomass in EBS pollock surveys indicate that historical changes in species and size/age composition were insufficient to introduce large uncertainties into the backscatter-tobiomass conversion in this survey. However, the use of historical relationships to convert backscatter to biomass assumes that conditions are in the range of those observed historically. In the case of the EBS USV survey, information from the fishery (Ianelli et al., 2020), recent surveys (Eisner et al., 2020), and the spatial and vertical distribution of backscatter observations all indicate that this was the case. However, one should not assume that this assumption will hold over long periods as species distributions change over time (Perry et al., 2005, Stevenson andLauth, 2019). Thus, the assumptions made in acoustic-only surveys should be periodically confirmed with biological sampling, for example by analysis of fishery data, or alternating USV-based and ship-based surveys.

Differences between ships and USVs
Fish are known to react to approaching survey vessels, typically biasing survey results lower (De Robertis and Handegard, 2013). Given their small size and lower acoustic and visual stimuli, USVs are likely to elicit fewer behavioral reactions than ships. Although this is an area where USVs are likely to perform better than ships, the potential effects of avoidance behavior must be considered so that changes in avoidance reactions affecting acoustic backscatter are not misinterpreted as changes in stock size. This is of particular concern for surveys of stocks prone to avoidance reactions, and/or distributed close to the surface where avoidance effects tend to be larger (Vabø et al., 2002). It is difficult to predict how fish will behave, but the impact of platform-specific reactions on survey estimates can be quantified by comparing paired measurements under survey conditions (De Robertis and Handegard, 2013). Comparisons of saildrones and the research vessel used in the EBS survey have shown that both platforms produce equivalent acoustic abundance estimates of pollock (De Robertis et al., 2019).
Hull-mounted echosounders are likely to be mounted shallower on small USVs than on ships. Although shallow transducers increase the detectability of near-surface organisms (Fernandes et al., 2003), they are more subject to signal attenuation from wind-swept bubbles at high sea states (Dalen and Løvik, 1981;Novarini and Bruno, 1982). The saildrone's transducers were mounted at 2 m, while the research vessel used in this survey is equipped with transducers on a drop keel (Knudsen, 2009) extended to 9 m. The potential for weather effects on the USV data was managed by identifying and removing pings with evidence of attenuation as well as pausing the survey at elevated wind speeds (De Robertis et al., 2019). Overall, attenuation of the acoustic signal is unlikely to be a major concern in this case as the survey was conducted in modest wind speeds and sea states. However, this could be a more serious limitation for USV surveys with shallower transducers (Swart et al., 2016) or conducted at higher sea states. Surveys in higher sea states may require deep-towed transducers (Greene et al., 2014), or sub-surface vehicles (Fernandes et al., 2003;Benoit-Bird et al., 2018).
USVs generally support small payloads and it is often necessary to instrument them with compact transducers. Small USVs cannot support the large 38 kHz narrow-beam transducers commonly used for fisheries surveys. Smaller transducers have wider beam widths, and will detect fewer fish near the seafloor. We applied a first-order correction for this effect (Mitson and Ona, 1996) which suggests that a relatively small proportion (3.2%) of the pollock detected by the research vessel were unavailable to the USV. The method assumes a flat seafloor over the beam width, normal incidence, and that fish abundance in the area above the ADZ is equivalent to that within it. In practice, these assumptions are likely to be violated to some degree, resulting in an under-estimate of fish abundance after correction. The survey results are likely robust to modest violations of these assumptions as much of the backscatter is well above the seafloor, and the EBS shelf is relatively shallow and flat (Zimmerman and Prescott, 2018). In cases where organisms are distributed close to the seafloor, at deeper depths, or present in steep or rugose habitat, the ADZ will be larger (Mitson and Ona, 1996) and the use of wider beam transducers may be a more serious limitation.

Future applications
This work demonstrates that USV deployments can be used to support fisheries management. Instrumented USVs suitable for acoustic surveys are commercially available (Verfuss et al., 2019). In this application, the saildrone vehicles and instrumentation used in this work were leased and the USV company conducted all USV operations and the echosounder calibrations. Given that USVs are increasingly accessible, their use is likely to become more widespread in acoustic survey applications. Before committing to this approach we encourage further analyses. One should quantify the financial cost of and tradeoffs associated with USV-based surveys in comparison to existing or alternate survey strategies. In addition, one should ensure that the resulting data products can fulfill management needs, for example via management strategy evaluation (Punt et al., 2016).
The lack of concurrent sampling of the species and size composition of acoustic scatterers remains the primary limitation to the use of autonomous vehicles in acoustic surveys. In the case of EBS pollock, we were able to covert backscatter to biomass estimates based on historical relationships, but this will not be possible in many instances. Given the current state of knowledge, USVs equipped with echosounders are likely to be most informative in well-characterized, low-diversity environments. Trawl estimates of species and size composition are critical for most acoustic surveys conducted for fisheries management (Simmonds and MacLennan, 2005). Alternative methods used to infer species and size compositions are, in their current state, insufficient to replace trawl sampling in most fisheries surveys. If USVs are to be used more broadly for fisheries surveys, improved methods to better characterize the species and size composition of acoustic scatterers must be developed.
Trawling is not likely to be possible from autonomous vehicles (Fernandes et al., 2003). Thus, the development of reliable methods to convert acoustic measurements to species and size without trawling has the potential to transform the use of echosounderequipped USVs. More widespread use of frequency response (Korneliussen, 2018), target-strength measurements (Levine et al., 2021), broadband measurements (Bassett et al., 2017), resonance scattering (Stanton et al., 2012), and application of machine learning (Brautaset et al., 2020;Proud et al., 2020), may ultimately result in better characterization of acoustic targets. Sampling methods that can be adapted to USVs such as small cameras (Fernandes et al., 2016), and environmental DNA samplers (Yamahara et al., 2019;Berger et al., 2020) may prove valuable. It may be useful to incorporate information from predator diets (Reiss et al., 2021) or fishery-dependent biological sampling (McClatchie et al., 2000;Chu et al., 2019), particularly if corrected for size and species selectivity (Millar and Fryer, 1999). Hybrid approaches in which ships and USVs work in tandem, with the ships trawling and the USVs increasing the density of acoustic measurements may prove effective. Survey situations differ, and improvements are likely to be incremental and situation-dependent. A pragmatic approach to select among the many possible approaches is to use simulation (Holmin, 2020) or re-analysis of existing data (Korneliussen, 2016;Proud et al., 2020) to quantify potential effectiveness.
This work, which we believe is the first USV-based survey to be used directly to manage a commercially important fish stock, should be viewed as an early attempt to use USVs to inform fisheries management. The backscatter-to-biomass unit conversion used here is a simple solution to a complex problem that may prove effective in other ecosystems where backscatter from a single species is dominant and historical surveys are available. Although the EBS pollock USV survey could not produce information on species, size, and age compositions typically collected from research vessels, it allowed the AT survey time series to be extended in a situation when gathering such data by other means was not possible.

Supplementary Data
Supplementary figures S1-S3 are available at the ICESJMS online version of the manuscript.