Use of the SmeltCam as an Efficient Fish-Sampling Alternative Within the San Francisco Estuary

Resource managers often rely on long-term monitoring surveys to detect trends in biological data. However, no survey gear is 100% efficient, and many sources of bias can be responsible for detecting or not detecting biological trends. The SmeltCam is an imaging apparatus developed as a potential sampling alternative to longterm trawling gear surveys within the San Francisco Estuary, California, to reduce handling stress on sensitive species like the Delta Smelt (Hypomesus transpacificus). Although believed to be a reliable alternative to closed cod-end trawling surveys, no formal test of sampling efficiency has been implemented using the SmeltCam. We used a paired deployment of the SmeltCam and a conventional closed cod-end trawl within the Napa River and San Pablo Bay, a Bayesian binomial N-mixture model, and data simulations to determine the sampling efficiency of both deployed gear types to capture a Delta Smelt surrogate (Northern Anchovy, Engraulis mordax) and to test potential bias in our modeling SFEWS Volume 19 | Issue 2 | Article 6 https://doi.org/10.15447/sfews.2021v19iss2art6 * Corresponding author: bhuntsman@usgs.gov 1 California Water Science Center US Geological Survey Sacramento, CA 95819 USA framework. We found that retention efficiency—a component of detection efficiency that estimates the probability a fish is retained by the gear, conditional on gear contact—was slightly higher using the SmeltCam (mean = 0.58) than the conventional trawl (mean = 0.47, Probability SmeltCam retention efficiency > trawl retention efficiency = 94%). We also found turbidity did not affect the SmeltCam’s retention efficiency, although total fish density during an individual tow improved the trawl’s retention efficiency. Simulations also showed the binomial model was accurate when model assumptions were met. Collectively, our results suggest the SmeltCam to be a reliable alternative to sampling with conventional trawling gear, but future tests are needed to confirm whether the SmeltCam is as reliable when applied to taxa other than Northern Anchovy over a greater range of conditions.


INTRODUCTION
Nearly all sampling methods in fisheries and wildlife are imperfect (Royle et al. 2005). Not detecting an individual during a survey does not

Use of the SmeltCam as an Efficient Fish-Sampling Alternative within the San Francisco Estuary
Brock M. Huntsman* 1 , Frederick Feyrer 1 , Matthew J. Young 1 exclusively mean a species is absent, but may reflect whether the species was available to or retained by the chosen gear. Such imperfections in the observation process propagate into the desired latent state metric (e.g., abundance, survival, diversity) that a study is designed to estimate. Although this source of bias can dramatically alter inference if unaccounted for (Kéry and Royle 2016), implementing appropriate sampling designs that can address catchability issues is not trivial.
In fisheries, the common catch equation explicitly defines the issue of detection efficiency (Catch = Detection Efficiency × Fishing Effort × Latent State Abundance, Walsh 1997). Detection efficiency can be further decomposed into components, which vary because of differing sources of sampling bias (Zhou et al. 2014;Hostetter et al. 2019). The common components of detection efficiency are the probability that a fish is present during sampling; the probability the fish, given it is present, is available to the gear; and the probability that a fish is retained once in contact with the gear, or retention probability (Hostetter et al. 2019;Vasilakopoulos et al. 2020). Various sampling designs have been developed to specifically target components of sampling efficiency (e.g., retention probability with covered cod-end trawling designs) or combinations of detection efficiency components (e.g., depletion surveys provide information specific to retention and availability probability, Hostetter et al. 2019). Often such sampling designs require replicate samples to be collected during demographic closure (i.e., when no movement, births, or deaths are occurring, Nichols et al. 2009;Amundson et al. 2014;Zhou et al. 2014), providing the appropriate sampling design to estimate bias in sampling efficiency. Designing experiments where closure is certain is relatively simple when surveys occur within small water bodies (e.g., using block nets within wadeable streams); however, it is difficult to ensure demographic closure in larger water bodies (large rivers, lakes, estuaries), and creative solutions are required. This difficulty is exacerbated for species with conservation status, especially if the most efficient gear types coincide with adverse effects (e.g., greater fish mortality when electrofishing or trawling).
The upper San Francisco Estuary (the estuary) provides municipal and agricultural water to satisfy most of California's freshwater demand, and is also home to multiple listed species that are dependent on this water supply. The Delta Smelt (Hypomesus transpacificus) is a federally threatened fish species (CDFW 2020) that is often the focus for resource managers tasked with balancing ecological needs and human demands because its decline is linked in part to freshwater outflows (Kimmerer and Rose 2018). A major concern with monitoring Delta Smelt populations is that it is sensitive to sampling mortality when collected by standard methods within the estuary (conventional trawls), and few sampling alternatives are currently available that can provide equally reliable indices of abundance (Feyrer et al. 2013). Recently, Feyrer et al. (2013) showed that an underwater video camera attached to trawling gear, referred to as the SmeltCam, could not only detect fish passing through the trawl but also reduced mortality by removing the need to handle fish (although some fish do get entangled in the mesh). The SmeltCam has been used to investigate the distribution of pelagic communities within different habitats of the estuary (Feyrer et al. 2017), suggesting its effectiveness to monitor taxa other than the fish species for which it was designed. Although current evidence points to the SmeltCam as a valid alternative to conventional trawling methodologies, no formal assessment of gear detection efficiency has been conducted.
Paired gear sampling designs can be used to assess detection efficiency among sampling gears in fisheries studies (Walsh 1997;Millar and Fryer 1999) and can be used to assess the relative efficiency of the SmeltCam as a sampling alternative to conventional trawling methods. For example , Coggins Jr. et al. (2014) combined baited traps and video cameras to estimate imperfect detection within an occupancy modeling framework. An important aspect of the paired gear sampling design is that the paired gear deployments simultaneously act as replicate, closed sampling passes (Coggins Jr. et al. 2014), a requirement for studies designed to estimate detection bias in sampling methods. Retention probability-the probability that an organism is retained by the sampling gear, conditional on it being available to the gear-is a component of detection efficiency commonly estimated for trawling gear deployed in the estuary. Paired gear studies have already been conducted within the estuary to estimate gear retention probability for Delta Smelt (Mitchell et al. 2017 and used to provide a less biased index of abundance throughout the estuary (Polansky et al. 2019). However, no such assessment has been performed for the SmeltCam, which is important if it is to be used as a viable alternative to conventional trawling surveys.
The goal of this study was to assess the gear retention efficiency of the SmeltCam relative to conventional trawling surveys deployed within the estuary. We used a paired gear sampling design, with the SmeltCam deployed simultaneously with a conventional trawl that targeted Northern Anchovy (Engraulis mordax) as a Delta Smelt surrogate to test the relative retention efficiency of the SmeltCam compared with conventional trawling methods. Our specific objectives were to (1) estimate how common environmental conditions during sampling might affect retention efficiency of the SmeltCam and a conventional trawl, (2) compare retention efficiency of the sampling gear types, and (3) use data simulations to determine potential bias in paired gear retention efficiency analyses.

MATERIALS AND METHODS Field Data Collection
We performed our study in September of 2016 within the Napa River and San Pablo Bay, California ( Figure 1). Surveys were conducted using the same equipment and operating protocols as normally deployed by the California Department of Fish and Wildlife (CDFW) during Fall Midwater Trawl (FMWT) surveys (Feyrer et al. 2007(Feyrer et al. , 2013https://www.dfg.ca.gov/delta/data), and near fixed sampling stations sampled by CDFW monitoring programs (Napa River stations 340 and 341; San Pablo Bay stations 310 and 311; https://www.dfg.ca.gov/delta/data/fmwt/stations. asp). The net is 3.66 m in width and height, and 17.6 m long, with nine tapered panels of stretch mesh that start at 14.7 cm near the mouth to 1.3 cm in the cod end.
The SmeltCam is a video camera apparatus (Genie HM1400, Teledyne Dalsa) attached to the cod end of a midwater trawl (see Feyrer et al. 2013 for more details about the SmeltCam's design). The purpose of the trawl-camera's design is to function as an open-ended trawling system in which fishes can be identified by trawling methods without the need to bring fish to the surface and retrieve them from a closed cod end (Feyrer et al. 2013). Unlike previous studies (Feyrer et al. 2013(Feyrer et al. , 2017, we included a closed cod end with the SmeltCam for our sampling design to test the fish retention efficiency of the SmeltCam relative to a typical oblique tow. Thus, our methods are similar to Coggins Jr. et al. (2014) in that the paired gear deployment served as replicate surveys in which gear retention efficiency could be assessed.
We conducted 65 paired SmeltCam and closed cod-end oblique tows over 4 sampling days, with each tow lasting roughly 12 minutes. All tows occurred against the current, 35 within the Napa River and 30 within San Pablo Bay. For analysis, we removed data points with missing covariate information, which reduced the number of tows to 52 (Napa River, n = 26; San Pablo Bay, n = 26). We conducted tows from 9/15/2016 to 9/16/2016 within the Napa River, and from 9/22/2016 to 9/23/2016 within San Pablo Bay, where the time between tows ranged from 17 minutes to 253 minutes. We estimated the volume of water filtered by the net using a mechanical flowmeter (model 2030R, General Oceanics) deployed off the side of the vessel during tows, with calculations provided by the manufacturer to account for tow speed and net size. We used a hand-held EXO2 YSI multi-parameter sonde (YSI Inc.) to determine water temperature (°C), turbidity (NTU), specific conductivity (μS cm -1 ), and salinity (ppt); and a boat-mounted sonar unit to estimate water depth (m) after tows were completed.
We counted Northern Anchovy separately for the SmeltCam and the cod end. For the SmeltCam, all images were processed by an expert in the lab, where fish were identified and counted. All fishes caught in the cod end were identified and counted in the field.

Statistical Analyses
We used a binomial N-mixture modeling approach (Royle 2004) to analyze the combined trawl-SmeltCam data to test for differences in retention efficiency between the two gear types. We used the dual gear binomial N-mixture modeling approach described by Graves et al. (2011) to analyze our data set, where two binomial N-mixture models were fit to count data collected by separate sampling methods within the same sampling area. This modeling approach assumes that latent state abundance (N) in contact with the gear was the same, regardless of sampling gear used, but differences in the observation process between gear types is the primary source of variability. We fit this same model with our dual gear surveys for Northern Anchovy with the following form: (1) (2) where and y is observed counts of Northern Anchovy from the SmeltCam (Camera) and the number of Northern Anchovy captured in the closed cod end of the trawl (Trawl) during each individual tow (i). Gear-specific retention efficiency (p) was constrained as a function of turbidity for the SmeltCam because of its presumed effect on video processing (DeCelles et al. 2017). The retention efficiency of the closed cod end was modeled as a function of the log-transformed total number of fish (plus 1 due to zeros) captured in the sample, because this is known to increase the sampling efficiency of trawl surveys because it blocks the mesh from which smaller fish may escape (Mitchell et al. 2017;Peterson and Barajas 2018). The latent state abundance parameter was also constrained by fixed tide effects ( i ; ebb, flood, high slack, low slack) and site effects (Napa River vs. San Pablo Bay), with the intercept ( N ) representing the ebb tide within the Napa River. We included the log transformation of the volume of water filtered during individual trawls as an offset. We also included a random effect ( ) for each observation in the observation process (p Camera i and p Trawl i ), and the latent state process (N i ) to account for over-dispersion (Kéry 2010;Harrison 2014), with as observation-level standard deviations.
We fit the dual gear N-mixture model using a Bayesian framework, with minimally informative priors assigned to all model parameters (Table 1). We conducted full Bayesian inference using Gibbs sampling (JAGS v4.3.0;Plummer et al. 2017) via Markov Chain Monte Carlo (MCMC) sampling with the jagsUI package in program R (Kellner 2019; R Core Team 2020). Final inferences for the N-mixture model were based on 6,000 samples for each parameter from the posterior distributions. Posteriors were generated from three independent chains, each with an initial length of 110,000 iterations, burn-in of 10,000 and a thinning of 50. We assessed model convergence by examining Gelman and Rubin convergence diagnostics ( R < 1.1; Gelman and Rubin 1992). We assessed "goodness-of-fit" for the N-mixture model by calculating a χ 2 discrepancy and a Bayesian p-value to assess model fit (Kéry and Royle 2016). Bayesian p-values indicate poor fit near values of 0 and 1, with the best fit at values near 0.5. Because the SmeltCam and the conventional trawl data were fit with two separate binomial N-mixture models integrated by a shared latent state abundance parameter, we assessed model fit by calculating a separate χ 2 discrepancy for the SmeltCam and conventional trawl observation processes (gear retention efficiency).
We used two criteria as evidence for Objectives 1 and 2. We designated parameters with 95% credible intervals that did not overlap 0 as evidence for environmental conditions that affected retention efficiency of both gear types and latent state abundance (Objective 1). Objective 2 was addressed in multiple steps. First, we predicted retention efficiency for each tow from both gear types. We took the mean from all tows, which resulted in a retention efficiency estimate for each gear type and posterior sample. We then compared retention efficiency predictions of the SmeltCam to closed cod-end midwater trawl predictions. We used the probability of direction to determine whether retention efficiency differences existed between the SmeltCam and the conventional trawl from posterior samples [Pr(Camera > Trawl)]; Makowski et al. 2019). If 95% of retention efficiency posterior samples were greater for one gear type than the other, we considered this evidence of an effect. Lastly, we used the probability of direction to also determine whether differences existed [Pr(Camera > Trawl)] based on the intercept estimates for each retention efficiency model ( Camera and Trawl ). We used this criterion because the intercept represents the expected mean of retention efficiency for each gear type when standardized covariates equal 0.

Simulations
We performed data simulations to test whether the model provided unbiased estimates of gear efficiency (Objective 3, see Appendix A for R code). We simulated data with linear constraints on the observation process (retention efficiency of the SmeltCam and trawl) and the latent state process (abundance), similar to the analytical framework described in Equations 1 and 2 but without the random observation effect, which was included to improve model fit for the observed data. The binomial model simulations included scenarios in which the intercept of both the SmeltCam and the trawl retention efficiencies were set at three different values (backtransformed from the logit scale = 0.4, 0.6, and 0.8) for a total of nine mean retention efficiency scenarios between the SmeltCam and trawl surveys.
The purpose of these simulation scenarios was to determine the extent that mean gear retention efficiency affected model reliability. We included a negative covariate effect (-1.0) for the SmeltCam simulation, which was the assumed effect of turbidity on SmeltCam retention efficiency. A positive effect also was included for the trawl retention efficiency simulation (1.0), an assumed effect of fish density on the retention efficiency of the trawling gear (Mitchell et al. 2017). A second simulation scenario included a range of sample sizes (or sites) to determine whether our sample size of 52 tows was sufficient to produce unbiased results of gear efficiency and how many tows (n = 25, 52, 88, 150) would be needed to Table 1 Model priors and posterior parameter estimates fit to the integrated binomial N-mixture model with SmeltCam and midwater trawl sampling gears in the San Francisco Estuary, California. Parameters reflect notation presented in Equations 1 and 2. Posterior values are means, with 95% credible intervals reported in parentheses. Bold numbers represent parameters with 95% credible intervals that did not overlap 0. The σ parameters could not overlap 0 because standard deviations can't be less than 0.

RESULTS
Northern Anchovy counts and habitat variables collected during trawling surveys varied between sampling locations. The highest number of Northern Anchovy captured with either gear type occurred in San Pablo Bay, where mean observed counts were approximately 8 to 12 times higher than counts from the Napa River (Table 2). Northern Anchovy counts were similar between gear types on each individual tow, although some variability was evident (Figure 2). Water-quality measurements also differed among sampling locations, with greater conductivity and salinity in San Pablo Bay, and all other measured water-quality variables higher in the Napa River (Table 2).

Objective 1
Analysis of the Northern Anchovy data with the integrated binomial N-mixture model showed adequate model fit (Bayesian p = 0.53 and 0.45 for the camera data and trawl data respectively, Table 1). No evidence was found that turbidity affected SmeltCam retention efficiency (Table 1), where 18% of the posterior samples for the turbidity effect were less than 0, and 95% credible intervals overlapped 0 (mean Turbidity = 0.30, 95% credible intervals = -0.34 to 1.09; Figure 3). Retention efficiency of the trawl increased with fish density (Table 1), with 98% of the posterior samples for the fish density effect greater than 0, and 95% credible intervals did not overlap 0 (mean Density = 0.63, 95% credible intervals = 0.07 to 1.33; Figure 3).

Objective 2
Multiple analyses estimated retention efficiency to be higher for the SmeltCam than the closed cod-end (trawl). Overall mean retention efficiency for the SmeltCam was higher than the closed cod end based on all 52 tows, with 94% of posterior samples (mean retention efficiency = 0.58, 95%  Figure 5). Similar results were found when we compared the retention efficiency intercepts, where 97% of posterior samples were higher for the SmeltCam (mean Camera = 0.44, 95% credible intervals = -0.46 to 1.65) than the closed cod end (mean Trawl = -1.25, 95% credible intervals = -2.99 to 0.60, Table 1).

Objective 3
Simulations demonstrated that relative bias in all parameter estimates decreased as the number of tows increased, retention efficiency of both gear types increased, and the number of replicated passes increased (Figure 6; see also Appendix B, Figures B1-B5). Patterns in RMSE for all parameters and simulation scenarios showed that RMSE decreased with higher sample size, number of replicates, and greater retention efficiency of each gear type (Figure 7). For simplicity, we only present results for intercept ( ) and slope ( ) parameters for simulation scenarios that matched our observed data set (SmeltCam retention efficiency = 0.6, and trawl closed cod end retention efficiency = 0.4; Figure 7).
Parameter interval coverage was high among all scenarios (all parameters were covered by 95% credible intervals in at least 83% of simulations); consequently, we similarly only present interval coverage of intercept and slope parameters for the same scenarios as RMSE (Figure 7). Simulation results for scenarios similar to our observed field data (number of tows = 52 and 1 replicate) showed coverage to be greater than 93% for the intercept and slope parameters, and RMSE to be low for both parameters (RMSE < 0.14, Figure 7).

DISCUSSION
We presented a combined gear sampling design within an integrated analytical framework to demonstrate the potential of the SmeltCam as a valid sampling alternative to conventional closed cod-end trawling gears used in the estuary. By allowing combined gear deployment to act as sampling replicates (see Coggins Jr. et al. 2014), we were able to estimate the retention efficiency of each gear type. Simulations further demonstrated that the integrated model can provide unbiased parameter estimates as long as model assumptions are met, and sample sizes are-as they were in the present study-high enough. The combined gear sampling design and analytical framework implemented here provides evidence that the SmeltCam is as efficient as a conventional midwater trawl at detecting fish available to the gear.
We found observed fish counts from the SmeltCam to be similar to trawl counts, but this direct relationship was variable, and reflected the imperfect retention efficiency of each gear type. (Again, retention efficiency is the probability that a fish is detected once it is in contact with the gear.) Retention efficiency was estimated for both gear types in this study, and is a common component of catchability estimated in fisheries studies (Walsh 1997). Mitchell et al. (2017Mitchell et al. ( , 2019 used covered cod end and paired gear sampling methods for midwater trawling surveys within the estuary. These surveys were not 100% efficient at capturing Delta Smelt, and smaller fishes were less likely to be retained than larger fish. Although we did not directly Figure 5 Posterior density of the mean retention efficiency using the combined SmeltCam and closed cod-end midwater trawl (Trawl) sampling gears analyzed with the integrated binomial N-mixture model test fish size during our surveys, fish size likely explains a considerable amount of variability in the retention efficiency estimates made for the trawl and the SmeltCam in our study. Fish size effects on retention efficiency estimates could be addressed with future modifications to the SmeltCam by incorporating multiple camera mounts (pairs of stereo-imaging cameras) that would allow for fish sizes to be determined (Rosen et al. 2013;Williams et al. 2015;DeCelles et al. 2017). The inclusion of covariates was important for our analyses because different aspects of sampling design can affect the efficiency by which gears perform. Turbidity was expected to negatively affect our ability to detect fish with the SmeltCam, similar to issues encountered in other image sampling approaches (DeCelles et al. 2017). However, we could not confirm whether there was a meaningful turbidity effect on the SmeltCam's performance. If turbidity is, in fact, important for SmeltCam performance, our analysis would indicate that our samples were not collected along a broad enough turbidity gradient (from 2.3 to 36.5 NTU) to detect an effect. We also included the total density of fish in the trawl as a factor affecting retention efficiency of the conventional trawling gear, and found evidence that the trawl was more efficient when fish densities were higher. Higher fish densities can fill mesh and prevent smaller fishes from escaping, a mechanism described for other trawling surveys in the estuary (Mitchell et al. 2017;Peterson and Barajas 2018).
The simultaneous deployment of the SmeltCam and closed cod-end midwater trawl was used to estimate gear retention efficiency for these gear types. Similar approaches have conceptually been taken in other studies, where cameras combined with other sampling gears (e.g., baited traps) are simultaneously deployed to estimate the same latent state variables (e.g., abundance, occupancy) using different gear types to inform the sampling efficiency of the complementary gear type (Bacheler et al. 2013;Kotwicki et al. 2013;Coggins Jr. et al. 2014). As expected, our simulations indicated greater accuracy and precision in parameter estimates when greater sample sizes and a greater number of closedreplicate passes were included (Zipkin et al. 2014;Duarte et al. 2018). However, increasing the number of true, closed replicate samples collected within the estuary to improve gear sampling efficiency is not a trivial matter, and cannot easily be addressed simply by increasing sampling effort. Confirming that sampling conditions are truly closed during replicate sampling passes is challenging, but is a requirement for estimating other decomposed forms of the detection process (e.g., the probability a fish is available to the sampling gear; Hostetter et al. 2019). For this study, we estimated the probability that a fish was retained by the sampling gear, but this value is conditional on a fish already being available to the sampling gear. Future efforts to reduce uncertainty in developing biological metrics from long-term trawling data sets in the estuary would greatly benefit from identifying sources of variability in detection efficiency beyond gear retention.
We used the binomial N-mixture model for our analysis, which can generate biased results if modeling assumptions are not met. First, the N-mixture model implicitly requires that double counting does not occur during data collection (Kéry 2018). The SmeltCam data could violate this assumption because a fish could move into or out of the scope of detection for the camera during a single tow. Similar counts between gears suggest that no strong violation of this assumption occurred during this study, but confirmation is not possible without unique fish identification. The N-mixture model also has issues with parameter identifiability when fit without covariate relationships in the observation or latent state processes ("intercept only models," Barker et al. 2018;Kéry 2018). However, we included covariate effects on both the observation and latent state processes, with strong covariate effects detected in the latter. Collectively, our results demonstrate that the SmeltCam provides abundance estimates that are as reliable as the conventional trawl, and can be used as a valid alternate sampling gear choice when monitoring at-risk populations such as the Delta Smelt.