Estimates of Predator Densities Using Mobile DIDSON Surveys: Implications for Survival of Central Valley Chinook Salmon

Abstract

The Sacramento–San Joaquin Delta (hereafter, “the Delta”) is one of the estuaries with the most invasive species in the world, and nonnative predators may be a major factor in the observed decline of Central Valley Chinook Salmon Oncorhynchus tshawytscha over recent decades. In order for managers to take actions that might reduce predation‐related mortality for these ecologically, culturally, and economically valuable fish, it is important to understand the factors influencing the distribution and abundance of piscivores in the Delta. In this study, we used a dual‐frequency identification sonar (i.e., DIDSON) to conduct mobile surveys to quantify the abundances of piscivores in the Delta. We then used these data to identify the habitat features that are correlated with the abundance of piscivores. Prior to conducting the surveys, we used DIDSON data from captured fish to develop an algorithm to distinguish piscivores from nonpiscivores with high confidence (98% accuracy). A generalized linear mixed‐effects model fit to these survey data indicated that predator abundances were most associated with areas of increased submerged aquatic vegetation patches, and channels that are straighter, with increased bathymetric complexity. When applied to the entire survey area, this model was successfully able to predict known areas of high predator densities. These results indicate that one approach to reduce predator densities in key locations throughout the Delta, and improve juvenile salmonid outmigration survival, is to reduce the extent of invasive submerged aquatic vegetation. Because experimental predator removals have been largely ineffective in the Delta, efforts to manipulate habitat to discourage nonnative predator recruitment and favor native species recruitment may provide a more effective solution to improve salmonid survival rates.

The Sacramento–San Joaquin Delta (hereafter, referred to as “the Delta”) is one of the estuaries with the most invasive species in the world and has undergone drastic changes in species composition over the past century (Cohen and Carlton 1998). Studies have shown that within the Delta, native fish species are often vastly outnumbered by nonnatives, with native species representing <0.5% of the total number of individuals captured (Feyrer and Healey 2003). Although it has long been suspected that nonnative predators in the Delta significantly affect the survival of native species, such as Chinook Salmon Oncorhynchus tshawytscha, empirical evidence linking salmonid mortality to predation is limited. A recent review of salmonid predation studies in the Delta suggests that the effect of piscivorous fishes on salmonid survival is among the most poorly understood factors affecting mortality, due in part to the difficulties in quantifying its effects (Grossman et al. 2013). Models based on current population trends, habitat conditions, and diet analyses suggest that Striped Bass Morone saxatilis alone could consume up to 29% of the emigrating salmonid population per year and impose a 28% chance of winter‐run Chinook Salmon extinction within 50 years (Lindley and Mohr 2003; Sabal et al. 2016). In order for managers to take actions that might reduce predation‐related mortality for these ecologically, culturally, and economically valuable native species, it is important to understand the factors influencing the distribution and abundance of piscivores in the Delta.

Understanding the characteristics that influence predator habitat preferences in the Delta is also important to inform restoration actions. Since the 1850s, the Delta ecosystem has been almost completely transformed from its original state due to land and water use management, ditching, agriculture, and flow modifications (Whipple et al. 2012). Millions of dollars are currently dedicated to habitat restoration actions in the Delta (Kondolf et al. 2008), but there is little information to guide managers towards what restoration actions will most effectively enhance native fish populations. This is particularly true because managers may be constrained by the physical and biological conditions of today's Delta, which has been heavily modified from historic conditions, and the uncertainty about how the Delta may be altered in the future with a changing climate (Whipple et al. 2012). Rather than restoring native fish populations, well‐intentioned restoration actions could inadvertently provide good rearing and foraging habitat for predators due to the lack of data on predator habitat preferences in the Delta. Thus, it is important to study what habitats predators most commonly utilize by surveying predator abundances in multiple habitats throughout the Delta.

It is difficult to collect representative spatial samples in a large area like the Delta using traditional survey methods (i.e., trawls, gill nets, and electroshocking) because these sampling methods are extremely time and labor intensive. In contrast, hydroacoustic methods are well known for their value in surveying large areas because they are less time consuming and labor intensive and are noninvasive (MacLennan and Simmonds 2013). While useful in surveying large areas, traditional acoustic methods, such as split‐beam sonar, make it difficult to distinguish targets and often increase uncertainty in estimates of target size, shape, count, and species. This is particularly true near scattering boundaries (e.g., dense vegetation or benthic structures that mask other nearby objects) or in horizontal applications where aspect angle has significant influence on the amount of sound reflected from the targets (Ona 1999; Horne 2000; Burwen et al. 2007). Multibeam imaging sonars, such as dual‐frequency identification sonar (DIDSON), vastly improve target resolution and the ability to observe behavior and measure the size, shape, and abundance of individual targets (Hateley and Gregory 2006; Xie et al. 2008; Martignac et al. 2015; Mora et al. 2015). The high quality of these images, combined with a high rate of capture (up to 21 frames per second), results in near video‐quality footage, which is why multibeam imaging sonars are often referred to as acoustic cameras. The ability to survey large areas with the DIDSON and to improve target resolution makes it a more efficient and effective sampling tool for sampling large areas, with complex bathymetry, compared with other hydroacoustic or traditional fish sampling methods.

In this study, we use a mobile DIDSON survey to assess the abundance of piscivorous fishes in the southern portion of the Delta. Our objectives were to (1) identify the major habitat features and environmental characteristics that are correlated with the abundance of piscivorous fishes and (2) quantify the abundance of piscivorous fishes in this region of the Delta during the primary period when juvenile salmonids are migrating. Achieving these objectives could provide managers with information to help understand the sources and locations of salmonid mortality and predation and provide insights that can be used towards habitat restoration in the Delta.

METHODS

Site selection

The methods we used to select survey sites are well documented in Michel et al. (2020) but are briefly summarized here. Although the entire Delta has over 1,100 km of waterways, our study focused on the lower San Joaquin River and South Delta (hereafter, “the South Delta”; Figure 1). This a region of extremely low survival for out‐migrating salmonids from the San Joaquin River drainage (Buchanan et al. 2013). This region is represented by low‐gradient, tidally forced rivers and sloughs that are largely channelized and leveed. All areas within the study area are freshwater but experience tidal influence and, during most years, flow in the upstream direction is observed during flood tides. At the southwestern end of the study site are two water export facilities that transport water from north to south. These export facilities transport up to 6.5 million acre‐feet of freshwater per year, most of which is used for agriculture, while the rest is used to provide, or supplement, drinking water for two‐thirds of California's population (Tanaka et al. 2011; Lund 2016). Multiple studies have shown a negative relationship between the amount of water exported from these facilities and juvenile salmonid survival (Kimmerer 2008; Tillotson et al. 2022).

We used generalized random tessellation stratified (GRTS) spatial sample selection (Stevens and Olsen 2004) to identify study sites. This method is a spatially balanced random sampling technique that ensures that all regions of interest are adequately sampled. The GRTS sample site selection method draws sites randomly, while balancing between different regions, and assigns them a draw number. The sample sites were selected to be spatially balanced across seven different regions of interest in order to best capture temporal trends throughout the South Delta. The seven different regions were selected based on differences in habitat characteristics (i.e., depth, width, and waterway type) and based on specific areas of management interest (Table 1). Franks Tract, Mildred Island, and the head of Old River are foci of management interest because they have been identified as predation “hot spots” (Grossman 2016). Furthermore, the upper Old River distributary connects to the water export facilities. We initially selected an overabundance of sites, and if any two sites were closer than three river kilometers (kilometers by way of the river) to each other, we randomly dropped one of those sites. A dropped site was replaced with a new site from an oversample list that maintained spatial balance. We selected 20 sample sites (Figure 1), 3 of which were visited every week (repeat sites) and 17 of which we visited only once. The three repeat sites (sites 1, 25, and 28) were included to differentiate between temporal and spatial variation in predation across the 6‐week sampling season. The repeat sites were selected to be spatially balanced across the seven different regions of interest in order to best capture temporal trends throughout the South Delta. One sampling site was visited on each day, and sampling was conducted from April 3 through May 13, 2017, to overlap with the primary outmigration season of San Joaquin River salmonid populations. A summary of the sampling dates and physical conditions at each sampling site is provided in Supplementary Table 1 (available in the online version of this article).

Survey methods

At each survey site, we demarcated a 1‐km reach that contained the GRTS‐drawn coordinates. Generally, reach end points were marked 500 m upstream and downstream of the GRTS location. However, reach end points were adjusted if they posed a potential safety risk. For example, if a reach end point included the entrance to a busy marina, we shifted the entire reach upstream so that the survey would not impede boat traffic. Within each 1‐km reach, we conducted longitudinal transect surveys using a 6‐m aluminum jet boat with two DIDSON units attached to the port and starboard sides (Figure 2). However, due to a unit malfunction, surveys on two dates (site 22 on April 29 and site 1 on May 1) were conducted with a single DIDSON camera mounted to the port side of the vessel. The DIDSON units were attached to the vessel with adjustable pole mounts, which allowed optimization of their pan, tilt, and depth (Enzenhofer and Cronkite 2005). We attached the mounts so the DIDSON was 1 m below the gunnel and approximately 30 cm below the surface of the water. We set the DIDSON range to a 10‐m window length (the maximum viewing window length in high‐frequency mode) to maximize the viewing range and image resolution. We started the viewing window 2.08 m from the lens to exclude areas immediately adjacent to the survey vessel because we expected fish to avoid areas close to the vessel (Figure 2). The DIDSONs were tilted approximately 10 degrees downward from horizontal to capture the upper water column just below the water's surface with minimal interference from the surface (Figure 2). Given the survey vessel was 2.5 m wide and the DIDSON settings allowed for a 12.08‐m viewing range (2.08‐m window start +10‐m window length), the vessel had a maximum survey width of 26.66 m (Figure 2). At most sites, we surveyed the entire channel, but at sites that were excessively wide, we restricted the survey width to approximately 180 m based on what we could reasonably survey within the limited survey time. Surveys were conducted during the last 2 h of daylight and the first hour after sunset, when predator and prey activity are typically high (Demetras et al. 2016). The boat tracks were offset approximately 12 m from the shore, or the edge of the survey area, to ensure that the full DIDSON beam reached near the shoreline. All field survey methods were approved by the Humboldt State University Institutional Animal Care and Use Committee prior to conducting field surveys (Institutional Animal Care and Use Committee Number 16/17.F.15‐A). See Loomis (2019) for further details regarding survey design and methods.

During surveys, an electric trolling motor propelled the boat to avoid disturbing fish along transects (Xie et al. 2008; Able et al. 2014) and to adjust the survey speed relative to currents. Surveys were completed at approximately 2 km/h, and survey speed was monitored via a Garmin global positioning system (GPS) installed on the survey vessel. This survey speed optimized video resolution and the number of transects that could be completed. At 2 km/h under normal weather conditions, a single 1‐km transect would take approximately 30 min to survey, allowing for approximately six 1‐km transects within the 3‐h survey window; however, poor survey conditions and/or technical difficulties often limited the number of transects that could be completed. Our goal during surveys was to repeat the same transect at least twice to have a replicate measurement of fish abundance and distribution.

Acoustic image analysis

The DIDSON survey images were analyzed using DIDSON software (V5.26.24; Sound Metrics Corporation, Bellevue, Washington). A single analyst reviewed footage manually at a rate of 1.5–5 times the rate of recording, depending on the complexity of the imagery. Any object within an image that was believed to be a fish was confirmed based on movement, profile, and image quality through multiple frames. The lengths of all fish were measured using the frame with the highest image quality and best profile. Ideal framing would align the fish perpendicular to the camera and within the center of the field of view to limit measurement error and bias (Hightower et al. 2013; Tušer et al. 2014). Length and thickness (the width of a fish from side to side) were both manually measured and recorded for all fish over 20 cm. We set 20 cm as the size threshold for potential predators of salmonid smolts based on published diet analysis of Delta predators (Nobriga and Feyrer 2007). Once image processing was complete, we applied additional species differentiation methods to distinguish between piscivores and nonpiscivores.

Species differentiation

In order to verify that we could distinguish between piscivorous and nonpiscivorous fish species in the Delta, we conducted DIDSON surveys in concert with two electrofishing surveys at the three repeat sites. The first electrofishing survey took place at the beginning of the study (April 11–13), and the second survey took place near the end of the study (May 9–11). We conducted DIDSON surveys at each site the evening prior to electrofishing and the following morning immediately before electrofishing. This allowed us to validate the fish community assemblage, compare the counts between electrofishing and DIDSON surveys, and collect predators of known size and species to build a DIDSON footage reference library.

Electrofishing surveys consisted of three single‐pass transects moving from downstream to upstream: one pass on each respective shore (or reach boundary) and one pass up the middle of the reach. Electrofishing was conducted by the California Department of Fish and Wildlife on an electrofishing boat with one boat driver and two netters. Any stunned fish visually estimated to be over 20 cm were netted, retained, and measured. Most fish were released at the end of electrofishing survey, but a subset of captured fishes was retained to be individually ensonified with the DIDSON to determine if acoustic data could be used for species differentiation. We tried to collect at least five fish of each species representing the range of sizes observed. Collected fish were identified to species, and we measured their fork length (mm), body width (mm), and body depth (mm).

Collected fish represented both piscivorous and nonpiscivorous groups of a variety of sizes and species. Two species of fish common to the Delta, Common Carp Cyprinus carpio and Sacramento Sucker Catostomus occidentalis, are nonpiscivorous with adult lengths that regularly exceed 20 cm in the Delta. We did not catch any Sacramento Sucker during the electrofishing surveys; thus, we could not quantify their abundance within the study region. Because the Sacramento Sucker are generally found in cool, clear waters and they have not been found in high abundances in the Delta (Moyle 2002), we assumed their abundances were negligible throughout the study region. Thus, the only nonpiscivorous species we collected for species differentiation was Common Carp since it was the only large (>20 cm) commonly occurring and nonpiscivorous species in this area of the Delta. We also collected individuals of the seven most prevalent predators from the electrofishing samples: Largemouth Bass Micropterus salmoides, Striped Bass, Redear Sunfish Lepomis microlophus, White Catfish Ameiurus catus, Black Crappie Pomoxis nigromaculatus, Brown Bullhead Ameiurus nebulosus, and Sacramento Pikeminnow Ptychocheilus grandis. A summary of the species collected during the electrofishing surveys is provided in Supplementary Table 2.

To observe the acoustic signatures of individual fish, we tethered fish and recorded them with the DIDSON from multiple distances. To tether the fish, we attached a jaw clip to a 1‐m monofilament line fixed to a horizontal pulley system deployed near the surface of the water. We then recorded DIDSON footage of the tethered fish at 1‐m increments ranging from 1 to 12 m, pausing for several minutes at each 1‐m increment to allow for natural fish movements. This created a reference footage library of known fish at several different ranges and orientations with movements like that of a free‐swimming fish.

We processed footage of tethered fish captured by electrofishing using Echoview (version 8.0.95, Echoview Software, Hobart, Tasmania) because of its ability to extract target strength (i.e., the acoustic signature of a fish) from DIDSON footage. Ensonified fish were first identified and delineated by an automated process. The software identifies objects (e.g., fish) based on the acoustic differences between pixels and their backgrounds and creates a layer of polygons delineating those objects. We then reviewed the processed footage for quality assurance. We visually inspected the footage and confirmed individual frames in which the delineation accurately represented the true size and shape of the tethered fish. We then exported the tabular data via individual target selection to ensure background objects were excluded and only the ensonified fish's data were exported. We used only the manually confirmed data for analysis.

We used a discriminant function analysis (DFA) to distinguish between piscivorous and nonpiscivorous fishes using DIDSON data. We used DFA to determine if it was feasible to distinguish between both (1) individual species and (2) piscivores and nonpiscivores. Using a combination of multiple identifying covariates, we created a set of candidate models and then used cross validation to identify the most appropriate model. In systems with high species richness, no single characteristic is likely to distinguish a species; thus, we tested multiple characteristics to differentiate species (Horne 2000; Able et al. 2014). Previous acoustic studies have used body length, height, length‐to‐height ratio, orientation, and target strength (i.e., acoustic backscatter from an individual target) to identify species with varying success (Ona 1999; Horne 2000; Mueller et al. 2010; Able et al. 2014; Martignac et al. 2015). We used similar metrics in our evaluation, but we used target width instead of target height because we wanted to distinguish the broad dorsal width of Common Carp. Prior to model fitting, we examined all metrics for collinearity and selected a single metric from any pair that had a correlation greater than 0.7. We then used cross validation to evaluate the performance of the model. Our cross validation procedure was to randomly divide the data into a training set (90% of the data) used to create a DFA and a validation set (the remaining 10% of the data) used to evaluate each model's prediction accuracy. We tested each model with 500 cross validation replicates and then used the mean of the resulting classification rates (i.e., the percent of fish correctly identified) to compare models.

Site density estimates

The density (D) of piscivores at each site was calculated using the following equation:

where $p¯j$ is the mean number of the piscivores observed per transect and $a¯j$ is the mean area surveyed per transect. Area surveyed per transect was calculated as the sum of the product of DIDSON viewing range and transect length for each DIDSON operated. We used bootstrapping to estimate a final predator density and quantify uncertainty. For each site, we randomly selected, with replacement, five predator density estimates based on all the transects conducted at the site. During the bootstrap selection, we equally sampled from the different locations (e.g., river left and river right) within site. For example, if three transects were conducted on each bank, we would have a total of six density estimates for that site to use in our bootstrap sample (Figure 3). To avoid bias from density differences on river left versus river right, we sampled equally from those two locations. We resampled our transects 1,000 times and used those values to calculate the mean and variance of density estimates for each site.

This estimate of piscivore density depends on multiple assumptions:

1. Each sample reach was closed to immigration and emigration during the 3‐h sample period.

2. Each observed fish was only counted once within a given transect (but that fish may be counted again in subsequent transects).

3. Fish observations are spatially dependent and transect spacing is sufficient to maintain the independence of nonoverlapping transects.

4. The survey volume was constant for any given frame. Since the cameras were focused on a fixed volume of the upper water column, we simplified calculations to a two‐dimensional surface.

5. Detection probability was constant and varied as a function of distance from the DIDSON cameras.

To account for the last assumption, each piscivore observation was weighted as a function of their range from the DIDSON camera. Because the viewing area and arc length of the survey window increase proportionally with range (arc length = 2π × radius × angle in radians), there is more area to view fish as range from the DIDSON increases (Figure 4). Conversely, at nearer ranges there is proportionally less area to observe associated predators. This calculation makes the simplifying assumption that the DIDSON field of view is two‐dimensional rather than three‐dimensional (assumption 4), which ignores the change in volume due to bathymetry in the field of view. Although changes in volume would affect the probability of observing fish, preliminary simulations suggested that the overall conclusions would not be considerably different. Thus, observed fish were weighted with a scalar calculated from the ratio of maximum survey window range (R_max) to the observed predator range (i) to account for the change in viewing area. Since the angle of the viewing window is constant, it is not necessary to include in the function. The weighting scalar applied to a predator fish at the observed range (w_i) is calculated as follows:

where R_max is the maximum viewing range for the transect and R_i is the observed range of a fish. As the observed range approaches the maximum range, the weight approaches 1. A fish observed at R_i = 2 m when R_max = 12 m would receive a weighting of approximately 2.3. This weighting function was tested with simulations using the WiSP R package (Zucchini et al. 2007). It was shown to provide accurate estimates of the true abundance and was robust to a variety of fish distributions and densities (Loomis 2019).

Regional density estimates

To identify the environmental factors influencing the abundance and distribution of piscivorous fish within the study region, we modeled predator densities at all the 1‐km sites. We chose a set of candidate predictor variables from the compiled data for the study region based on their hypothesized influence on predator distribution. Michel et al. (2020) provides a detailed description of how these data were collected. The environmental covariates that we included were average depth (m), depth coefficient of variation (depth CV), channel sinuosity, water velocity (m/s), turbidity (NTU), and temperature (°C). Channel sinuosity was calculated by dividing the total length of the channel (i.e., river kilometers) by the straight line distance from the beginning to end of each site. Each of these factors were hypothesized to either directly affect habitat suitability, fish metabolism, and seasonal movements or affect a predator's ability to effectively forage in estuarine and riverine environments (reviewed by Čada et al. 1997; Gregory and Levings 1998; Sweka and Hartman 2003; USEPA 2009; Callihan et al. 2014). We also tested the effects of two types of vegetation on predator density: native vegetation (tule Schoenoplectus acutus patches) and invasive vegetation (submerged aquatic vegetation [SAV]). Predators are often associated with emergent and submerged vegetation, and it has been hypothesized that the spread of invasive plants, which dominate the SAV community, has contributed to the success of invasive predator species (Nobriga and Feyrer 2007; Conrad et al. 2016). We used two metrics of SAV as covariates in our models: (1) total patch area (m²) and (2) patch density (patches/100 m²). Finally, there is growing concern over the effects of anthropogenic alterations, such as flow diversions, on the interactions between salmonids and piscivores (Feyrer and Healey 2003; Sabal et al. 2016). To represent anthropogenic alterations in our model, we included covariates for both the area (m²) and count of human‐made structures and the length of levees along the channel (m).

We fit mixed‐effects linear models and used model selection to identify which combination of model covariates was the most parsimonious. Our response variable in this model was the log‐transformed predator density in the 1‐km reach. All area‐dependent covariates were first scaled by the area of their respective study site as proportions. All continuous variables were standardized (i.e., z‐transformed) by subtracting the means and dividing by the standard deviation. Prior to model fitting, we conducted pairwise correlations among covariates to assess collinearity and removed a single covariate from any pair with a correlation >0.7. We fit mixed‐effects linear models to the data using the glmmTMB package (Brooks et al. 2017). To account for the repeated measures of sites 01, 25, and 28 throughout the 6‐week study period, we included a random intercept for week in all candidate models. Using a two‐step process, we performed model selection using Akaike information criterion corrected for small sample size (AIC_c) and cross validation. Because this was an exploratory analysis, our candidate model list consisted of all possible subsets of five or fewer predictor variables. We ranked these models based on their respective AIC_c scores. For each model with a delta AIC_c score of two or less, we then used leave‐one‐out cross validation and calculated the r² between the observed and model‐predicted values to provide a relative estimate of goodness of fit.

The best model based on both delta AIC_c and cross validation was then used to extrapolate measured predator densities within all 1‐km reaches throughout the study region. For each 1‐km reach, we calculated sinuosity and depth CV using the same methods and data sources as for model construction (Michel et al. 2020). Because we did not have measured extents of SAV outside of the study areas, we used remotely sensed SAV data collected by researchers at the University of California–Davis (Hestir et al. 2008). Of the available SAV data, the 2015 data set had the most extensive coverage and was thus used for extrapolation. Due to the low resolution of the survey equipment used in the 2017 environmental data collection, features <5 m across their longest axis were excluded from SAV delineation (Michel et al. 2020). To maintain consistency with this methodology, the 2015 data set was filtered to exclude SAV polygons whose longest axis was shorter than 5 m. The 2015 data also included SAV polygons occurring in water deeper than SAV is typically found, which is unlikely. Thus, we excluded SAV from water deeper than 5 m (Shruti Khanna, University of California–Davis, personal commnication). Sinuosity and depth CV were then standardized (z‐transformed) to the data used in model construction. We scaled the SAV data both by the area of each 1‐km reach and then by their own mean and standard deviation. These data were then used to predict predator densities across the study region using the results of the best model.

RESULTS

Species Differentiation

The two electrofishing surveys at the three repeat sites captured a large number of nonnative piscivorous fish. The electrofishing surveys took place at site 1 (April 11 and May 11, 2017), site 25 (April 12 and May 9, 2017), and site 28 (April 13 and May 10, 2017). A total of 624 fish over 20 cm were captured, measured, and released (Supplementary Table 2). Electrofishing catch compositions were typically dominated by Largemouth Bass followed by Striped Bass and sunfishes (Figure 5). Striped Bass were notably abundant at site 1 on May 11, 2017, and absent from catches on both sampling occasions at site 28. All other combined species typically composed <15% of the total catch. With the exception of the variability in Striped Bass abundances, catches were relatively stable between these two sampling events.

There was no relationship between the electrofishing catches and the DIDSON abundance estimates collected on the same mornings from the same transects. Linear regression indicated poor correlation between electrofishing catches and DIDSON observations of fish greater than 20 cm in length (r² = 0.005, P = 0.888) (Supplementary Figure 1 [available in the online version of this article]). Electrofishing minutes were only recorded during the May sampling events, so we did not have a sufficient sample size to use catch per unit effort as a response variable. We observed much less variability in the number of fish captured by electrofishing than observed with the DIDSON on different dates at the same site. For example, with the DIDSON we observed a difference of nearly 100 fish at site 25 between weeks 15 and 19. In contrast, the electrofishing catches were comparable (week 15 = 92 and week 19 = 72) at each site on the two different dates. We observed a similar discrepancy between DIDSON counts and electrofishing catches at site 1. Comparison of catch data alone does not account for the inherent difference in the two methodologies, and a more rigorous study would be required to accurately compare these two methods; however, these initial findings suggest the two different survey methods would lead to very different conclusions about fish abundances.

We tested the accuracy of species discrimination using DFA with 2,248 acoustic measurements from 42 unique fish sampled during electrofishing efforts (Table 2).

Discriminant function analysis was effective in discerning Common Carp from piscivorous species but provided poor confidence in distinguishing individual species from each other (Figure 6). We excluded fish width as a model covariate due to collinearity with fish length, but we still included the length‐to‐width ratio. Thus, we ended up with 15 candidate models for species discrimination. The same 15 candidate models were used for piscivores versus nonpiscivores. The DFA models determining individual species using combinations of target strength, orientation, length, and length‐to‐width ratio had classification accuracy between 14% and 37% (Table 3). Models were more accurate at distinguishing piscivores from nonpiscivores, with classification accuracy ranging from 63% to 98% (Table 3). All DFA models used to distinguish piscivores from nonpiscivores that included length had a classification accuracy of at least 97%. Although the model that used length as the only coefficient was highly accurate, we decided to use the model that included both length and the length‐to‐width ratio because of the apparent differences between groups in all the morphometrics used and previous success with a similar model (Mark Bowen, presentation at the 10th Biennial Bay–Delta Science Conference, 2018). We note that much of the separation achieved by DFA was likely due to the differences in the size distribution between the samples. All of the nonpiscivores (i.e., Common Carp) in the electrofishing sample were more than 50 cm, and the majority of the piscivores were smaller (Table 2). However, we believe these sizes were representative of the fish population in this area of the Delta. Applying the final function to the DIDSON survey data resulted in 6,434 observations classified as piscivores and 186 classified as Common Carp.

Site Density Estimates

Throughout the 6‐week survey period, we sampled 227 transects and collected approximately 193 h of DIDSON footage. This footage yielded a total count of 6,638 fish over 20 cm after image analysis. Mean density estimates from the 35 sampling days ranged from 7.34 to 56.99 predators/100 m² (mean = 18.9 predators/100 m², standard deviation = 10.85) (Figure 7). There were two sites with outlying density estimates greater than two standard deviations above the pooled mean and only a single estimate occurring below one standard deviation of the pooled mean. The highest estimated predator density occurred at site 25 on May 8, likely the result of an immigration of Striped Bass to the area (Figure 5). A similar movement of Striped Bass may have been responsible for the third highest density occurring on May 2 at site 25, but this cannot be confirmed because we did not conduct any catch sampling on that date. The lowest predator density estimate occurred at site 19 on April 29, a site along Old River near the entrance to Discovery Bay, a popular marina and waterfront housing development. There was no indication of any temporal trend in piscivore densities throughout the study area or in any of the repeat sites (Figure 7).

Regional Density Estimates

Results from the model selection process indicated that the most parsimonious mixed‐effects linear model included the following covariates: sinuosity, SAV patch count, and depth CV. Our results indicated there was no support for week as a random effect, but we kept it in the model because that was how we planned to analyze the data a priori based on our survey design. All five of the candidate models with a delta AIC <2 included the covariates SAV patch count and sinuosity. In addition, three out of the five models contained depth CV. An important distinction is that while both SAV patch count and total SAV area were included in the global model, only SAV patch count appeared in the top models. This indicates that predators may be more likely to select habitats with a patchy distribution of SAV as opposed to large, dense mats. Interestingly, there was little support for the physical covariates that are often used to described fish habitat (e.g., temperature and dissolved oxygen). The most parsimonious model included only the three covariates that were in most of the top five models (Table 4). Based on leave‐one‐out cross validation, this model also had considerably more predictive power (multiple r² = 0.12) than a simpler model with just SAV patch count and sinuosity (multiple r² = 0.04); thus, the model with depth CV was chosen for predicting predator density at a regional scale. This model indicated that the highest densities of piscivores would be found at sites that are less sinuous, are bathymetrically complex, and have many patches of SAV (Figure 8; Table 5). Extrapolation of the regional‐scale model resulted in predicted predator densities ranging from 6.81 to 329.62 predators/100 m², with a mean of 20.79 predators/100 m² (Figure 9A). Noteworthy predictions include many high‐density reaches along the upper main‐stem San Joaquin River, including the highest predicted predator density at the head of Old River (Figure 9B). This is an area of management interest because it is where fish can remain within the main‐stem San Joaquin River or enter the interior Delta, where they can become entrained in the water export facilities (Monsen et al. 2007; Cavallo et al. 2011; Buchanan et al. 2013), and has been found to be a potential predator hot spot in past research (Vogel 2010).

DISCUSSION

The major conclusion from our study was that we were able to estimate densities of piscivores throughout our entire study area using our mobile DIDSON survey approach. We found that the DIDSON surveys were fast and efficient ways to survey the abundances of fishes within the channel. The major current issues with using DIDSON to conduct these types of surveys are that (1) species identification is not feasible and (2) data processing can be time intensive and onerous. Species differentiation has long been a major objective of acoustic surveys, with extremely limited success (Horne 2000; Martignac et al. 2015). As with previous studies, we had limited success differentiating between individual species, but that was not necessary to achieve the goals of our study. We only needed to be able to identify, and quantify, potential predators of juvenile salmonids. Using images of known fish, we were able to distinguish between piscivores and nonpiscivores with near perfect (98%) accuracy. Our ability to discriminate between piscivores and nonpiscivores relied on large morphometric differences between those groups, which will not occur in most ecosystems. However, we believe there is potential to use a combination of metrics (e.g., target strength, length, habitat preferences) to improve species differentiation capabilities. With regards to data processing, we are hopeful that this will become more automated as image processing and machine learning capabilities improve (Purser et al. 2009; Gorsky et al. 2010). Regardless, our study has demonstrated that this approach can be used to understand the covariates that influence predator densities and to identify the locations that predators are expected to be the most abundant. Both of these capabilities could be invaluable to managers seeking to mitigate the influence of these nonnative predators on native species.

The DIDSON sonar produced relatively high‐quality images compared with other acoustic devices, and we were able to use these images to develop high‐quality piscivore density estimates. Previous research that studied the effect of Delta predators on juvenile salmonids in the same area of the Delta used traditional hydroacoustic survey methods to quantify piscivore densities (Hayes et al. 2017). That research combined acoustic surveys of predator densities with relative predation rates measured by predation event recorders (PERs; Demetras et al. 2016). A PER is a free‐drifting buoy that can track when and where a live tethered smolt is predated via a magnetic switch, timer, and GPS device. Results from Hayes et al. (2017) indicated that it was difficult to correlate acoustic survey data with predation rate estimates from the PER experiments, most likely because of the asynchrony in the timing of the surveys and because the traditional hydroacoustic surveys could not adequately distinguish larger predators from smaller fish and nonpiscivores. The DIDSON was able to address both of these problems. Because of our increased resolution, we could estimate where predators were located while a concurrent PER study was conducted by Michel et al. (2020). Our high‐quality data combined with PER data from the Michel et al. (2020) study allowed us to conclude that the mean spatial–temporal distance between a piscivore and a PER was correlated with predation rates.

We also observed that our DIDSON counts did not agree with the electrofishing counts conducted at the same site within hours. There are a few potential explanations for the lack of agreement between these two survey methods. The first is that we did not have any estimate of effort for the electrofishing surveys and, therefore, could not compare the catch per unit effort between the two methods. This standardization might have resulted in more similar metric between the two methods. Another possibility is that the fish we observed with the DIDSON were not present when the electrofishing survey took place. We do not believe this is a likely explanation because, as we discuss below, none of the most important covariates that predicted predator density varied with time. Another explanation for the lack of agreement is that the counts from either, or both, of the survey methods were not accurate for some reason. We suspect that it is most likely that the electrofishing counts were not accurate for multiple reasons. First, the electrofishing surveys were not depletion surveys and, therefore, have no estimate of capture efficiencies and could have some size bias to the catch (Odenkirk and Smith 2005). Secondly, we suspect that the electrofishing survey did not accurately sample all species in the community. Previous research has found differences in boat electrofishing capture probabilities for different species depending on size, habitat preferences, and behavior (Ruetz et al. 2007; Reid et al. 2021). Thus, it is possible that the electrofishing surveys at these locations only provided an accurate sample for a subset of the fish community.

Our regional predator density model indicated that predators were found in increasing densities in habitats with more patches of SAV. Largemouth Bass, the most abundant predator captured during electrofishing efforts, are known to reside and forage within areas of dense vegetation (Savino and Stein 1989). Largemouth Bass are also known to seek submerged cover, including SAV, to search out prey that may be taking refuge but also use the structure as cover for ambush predation (Wanjala et al. 1986; Savino and Stein 1989). Furthermore, the proliferation of invasive SAV in the Delta has been linked to the expansion of Largemouth Bass habitat (Brown and Michniuk 2007; Conrad et al. 2016). It is likely that naïve salmonids, which did not evolve in a landscape where ambush predators lurk among the vegetation, have not yet developed defense mechanisms to avoid and escape these types of attacks.

Regional distributions of predators were also mediated by sinuosity and coefficient of variation of depth. We found a positive relationship between predator density and the depth CV, suggesting that predators are selecting habitats with more abiotically complex structure, likely for the same reasons that they are selecting areas with more complex SAV structure. Contradictory to this notion, the negative relationship observed with sinuosity suggests that predators are selecting more linear sites, which are generally associated with lower structural complexity and habitat heterogeneity. It is possible that sinuosity is collinear with unmeasured habitat variables such as substrate type. Many linear channels of the Delta are the result of levee construction, which may be armored with riprap. Based on information posted on several websites and discussion boards viewed by the authors, recreational anglers may target these riprapped embankments due to encounters with Largemouth Bass and other centrarchid fishes. Brown and Michniuk (2007) also speculated that the replacement of natural banks with riprap material could explain the dominance of nonnative fish in littoral habitats of the Delta.

Despite the value of our study, there are some important caveats to consider in interpreting the results. First, the data were collected during a single, unusually wet spring (CDWR 2018). Therefore, the relationships we observed may be unique to the unusually high flows and cool temperatures observed in the spring of 2017. Secondly, to simplify our calculations, we did not account for the volume of water surveyed, which could affect the probability of observing a fish within the DIDSON beam when the river bottom was in the field of view. In our case, we believe that any error due to this simplification would be minimal and would not affect our overall conclusions. However, this potential problem may be something to consider in future studies. Finally, the regional model was developed based on the cumulative behavior of at least 12 different predators. While many of the centrarchids exhibit similar habitat selection of littoral habitats, Striped Bass and White Catfish typically select mid‐channel, open‐water habitats (Feyrer and Healey 2003; Michel et al. 2018). These differences in habitat preferences may partially explain why there were no time‐varying physical covariates in our model, such as temperature and dissolved oxygen, and why the overall model goodness of fit was low (r² = 0.14). Because centrarchids numerically dominated the catch composition during electrofishing efforts, the observed relationships are likely more representative of the choices these species make, and the uncertainty of these models is due, in part, to the observations of the typically less abundant midchannel‐dwelling species. Furthermore, Striped Bass were transient throughout most sites and could either dominant the catch or be relatively absent. It is currently unclear what factors drive their movement patterns through the Delta.

Species differentiation to a finer resolution than what we could achieve would greatly improve the predictive power of these models. This may be possible with a larger sample size of predators from each species and size‐class, since preliminary results indicated a strong potential to discriminate some of the well‐represented species based on target strength and orientation (Supplementary Figure 2). We also suspect that our current species differentiation function is biased and would benefit greatly from a larger sample size. All of the Common Carp recorded in reference footage were over 56 cm, while all predator species used were <50 cm. The DFA models trained using only fish length from “predator” and “carp” groups also resulted in extremely high classification rates (>97%), suggesting that the other measured morphometric values were not contributing substantially to the discrimination function. Predator fish in the Delta do reach sizes over 50 cm and Common Carp also occurs in smaller sizes, but we did not capture any in the electrofishing sample. Without access to an additional database of morphometrics and acoustic measurements for these species, we were unable to refine the DFA further and we assume that the sampled fish were representative of the fish population present in the study area during the survey period.

Management Implications and Future Research

These model predictions indicate that predator densities in the upper San Joaquin River between the Stockton ship yard and the head of Old River were consistently high (mean = 32.04 predators/m²), which could help explain why survival estimates are typically very low through this region (Buchanan et al. 2018). Based on our results, one option that managers could consider to reduce predator densities along these migration corridors is to reduce the extent of invasive SAV, such as Brazilian waterweed Egeria densa. Because experimental predator removals have been largely ineffective in the Delta, efforts to manipulate habitat to discourage nonnative predator recruitment and favor native species recruitment may be more effective (reviewed in Bowen, presentation; J. D. Wikert, presentation at the 10th Biennial Bay–Delta Conference, 2018). Model predictions using the landscape model developed in this study indicate that a reduction of SAV patches along the upper San Joaquin River between the head of Old River and Stockton by only 50% could reduce predator densities by approximately 18%, while a complete eradication of SAV in this region could reduce predator densities by approximately 32%. Invasive SAV, including Brazilian waterweed, has been identified as problematic throughout large portions of the San Joaquin River and Delta due to interference with water conveyance and recreational and commercial boat passage and threats to natural ecological processes (Moran et al. 2021). Annual weed control efforts are undertaken to mitigate these effects but are limited by high costs, lack of funding, a complex regulatory structure, and a lack of consistent monitoring (Ta et al. 2017). Targeting specific areas and migration routes (e.g., the main‐stem San Joaquin River) could reduce total mitigation costs and potentially have a large impact on reducing predation‐related mortality of juvenile salmonids. To determine if a vegetation removal program is effective at reducing predation‐related mortality, it could also be valuable for managers to develop an adaptive management plan to evaluate such a program (Walters 1986; Williams 2011).

ACKNOWLEDGMENTS

Funding for this project came from the California Department of Fish and Wildlife (CDFW). The two DIDSON units were also loaned to us by the CDFW. We also thank numerous individuals at the CDFW for their assistance in conducting the electrofishing surveys. We are extremely grateful to everyone that volunteered their time to assist with fieldwork to make this project possible. We thank Andre Buchheister, Darren Ward, Ethan Mora, and Brian Burke for providing extremely valuable comments and edits on early drafts of this manuscript. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. At the time of publication, the data was not publically available through the CDFW. There is no conflict of interest declared in this article.

REFERENCES