Evaluating Batch Imaging as a Method for Non-Lethal Identification of Freshwater Fishes

Abstract

Freshwater fish community surveys are an important component of aquatic ecosystem management. However, the standard method for taxonomic identification currently used for these surveys, wherein fishes are manually identified in the field by a taxonomic expert, has several shortcomings. These include handling-related fish injury and mortality, the need for a fish-identification expert to be present during field sampling, and additional fish mortality due to physical voucher collection. These shortcomings may be overcome using new methods such as environmental DNA (eDNA) or image analyses. While eDNA can provide fish community data through metabarcoding, it is costly and provides little ecological information. A novel, image-based method for taxonomic identification (“batch-image identification”), which addresses the shortcomings of standard and eDNA methods, was tested in this study. Fishes were captured in the field and photographed in small groups (“batches”) within fish viewers for subsequent identification by taxonomic experts. Comparing taxonomist-based identifications from batch images to specimen-based identification, batch-image identification yielded an overall species-level correct-identification rate (CIR) of 49.7%, and an overall genus-level CIR of 61.2%. CIR increased with taxonomist expertise, reaching 83% when identification was performed by expert taxonomists. Batch-image identification data also produced rarefaction curves and fish-length measurements comparable to those obtained through standard methods. Potential methodological improvements to batch-image identification, including procedural adjustments and alternative identification methods, provide direction for the continued testing and improvement of this method.

1. Introduction

Surveys of freshwater fish communities are routinely undertaken globally for a variety of purposes, such as assessing geographic distribution [1], conducting biological monitoring [2,3,4], and informing conservation [5] and natural resource management [6]. As a result of their wide range of applications, freshwater fish community surveys are frequently conducted, and their methods are continuously being made more efficient, effective, and safe.

Currently, identification of fishes during freshwater fish surveys typically takes place in the field, which requires fish to be removed from the water and handled manually, such that the morphological features required for identification may be examined (e.g., [7]). Fishes are then released; however, a representative sample of physical or digital specimens (“vouchers”) may be euthanized (for physical vouchers) and retained, as a method to verify the identity of the fishes collected [7,8]. This current practice of manual, field-based identification has several shortcomings. Because fishes often need to be identified and released in the field, at least one individual on the sampling crew must be proficient at fish identification. The need for a fish-identification expert on each field crew limits the sampling that can be conducted, as there are limited numbers of fish-identification experts as a result of the recent trend of fewer biologists being trained and employed in fish taxonomy [9,10]. Manual identification of fishes in the field requires additional time to be spent at each sample site. Identifying and enumerating fishes by hand also requires extensive handling, which can inflict stress and physical harm [8]. Injury and stress can lead to an increased risk of post-release mortality for sampled fishes [11,12,13,14], which is particularly problematic when sampling is being conducted on imperiled populations [15]. The requirement by standard protocols for voucher specimens to be retained [7,8] is an additional source of fish mortality and requires the expenditure of materials and time for transport and maintenance of preserved specimens.

To overcome these shortcomings, alternative fish identification methods can be utilized in freshwater fish surveys. For example, the use of environmental DNA (eDNA) can determine the presence [16], and potentially the abundance [17], of fishes without having to physically capture and handle them, improving survey efficiency. However, it is unlikely that eDNA will completely replace conventional sampling, as eDNA methods can be significantly more costly compared to existing survey methods while providing less ecological information [18,19,20], and there is still a need to capture specimens for a variety of reasons, such as: (1) confirming identity, as there may be false presences and absences in eDNA results for a variety of reasons [21]; (2) confirming abundance, as further research is required before eDNA alone can produce reliable abundance estimates [17]; and (3) measuring lengths to determine length–frequency relationships for stage-based fisheries management [22]. Therefore, it is important to continue to develop additional methods to improve the process of taxonomically identifying fishes.

Image-based techniques are another method for taxonomic identification with the potential to be employed as an alternative to standard manual identification. Instead of manually examining fishes in the field for identification, fishes can instead be imaged (i.e., photographed or videographed) in the field and subsequently identified from the images at a later time. Identification of fishes from images can be performed by taxonomic experts or, with recent advances in computer vision and artificial intelligence (AI), can also be conducted through automated approaches. Numerous recent studies have employed computer vision and AI to identify fishes in both photo and video imagery and have achieved high rates of accuracy (e.g., [23,24,25]). However, there remain obstacles inhibiting the utility of AI-based fish identification tools, including the requirement for a high level of consistency in imagery for identifications to be accurate, limited taxonomic applicability of algorithms (usually a single or small set of species or genera), and difficulties with identification when images contain multiple fishes [26]. As a result, image-based fish identification by taxonomic experts remains relevant despite the advent of automated approaches.

Image-based identification of fishes has numerous advantages compared to manual identification, including: (1) images allow fishes to be identified after sampling, minimizing the need for experienced fish taxonomists in the field, or, in the case of AI, eliminating the need for conventional sampling and taxonomists [26]; (2) fishes either do not need to be handled at all (when using in situ imagery; e.g., [27]) or only need to be handled when transferred into the photography setup (i.e., fish viewer) after capture, allowing for less handling than in manual identification methods and minimizing fishes’ time spent out of the water, improving fish welfare [28,29]; and (3) images can provide digital vouchers of captured species [7], reducing the need to euthanize and store fishes in order for physical vouchers to be retained. However, there remain some disadvantages related to the use of imagery to taxonomically identify fishes. The effectiveness of identification based on in situ imagery is inhibited by a lack of control over in situ conditions (e.g., lighting, turbidity), which can result in poor-quality images [23,24,25,27,30] and also by the difficulty of obtaining a representative sample of the fish community due to issues of sampling biases [30,31,32]. Conversely, obtaining imagery of captured fishes in the field can present logistical challenges, with complex arrays of expensive equipment often being required to obtain images of a high enough quality to allow for taxonomic identification, and additional time being required for each image taken [23,24,25,33,34].

In this study, we investigated a novel method of photo-based fish identification that aims to address some of the limitations of existing methods. Our method, which we have termed “batch-image identification”, is based on photographs of groups of live fishes in fish viewers (narrow fish tanks) in the field. Instead of imaging one individual at a time, captured fishes are placed in a fish viewer and photographed in small groups (“batches”), allowing for more time-efficient collection of photographs and reduced handling time for each individual fish. By utilizing a relatively simple and inexpensive set of equipment, this method also minimizes constraints presented by setup time and budgetary concerns. To our knowledge, our study is the first to explore the use of batch images to taxonomically identify freshwater fishes. To assess the accuracy and feasibility of batch-image identification, we compared batch-image identifications conducted by experts to laboratory identifications of the same fish samples. We analyzed the rate at which fishes were correctly identified based on the images when compared to laboratory identifications (correct-identification rate [CIR]) and the factors that influenced this rate. As fish length is often an important demographic metric accompanying taxonomic identification during fish surveys [35], we also evaluated the accuracy of measuring the length of specimens on the images compared to measuring them in the laboratory. Additionally, to assess the potential of using batch-image data to calculate species richness (a common objective of biodiversity surveys), we compared species richness estimates and associated rarefaction curves calculated from batch image and laboratory datasets. We hypothesized that CIR would not be 100% because some identification characters would not be visible in the images and, as a result, we hypothesized that species richness estimates would be lower for the batch image dataset. We hypothesized that there would be no significant difference in fish lengths because the images were taken in a narrow aquarium with minimal depth distortion.

2. Materials and Methods

2.1. Overview

We evaluated the utility of batch-image identification through a field study and subsequent statistical analyses. First, we undertook field sampling to obtain physical specimens and batch images (Figure 1) for the study. We then identified the physical specimens to the species level in the laboratory, creating a control identification dataset. To empirically analyze the effectiveness of batch imaging as an identification method, we created an online survey in which participants with knowledge of freshwater fish taxonomy were asked to identify the fishes in the batch images obtained from this study. The data obtained from the survey were used to create a dataset of batch-image-based fish identifications that was compared to the laboratory identification dataset. We then performed statistical analyses on the data obtained from the survey to test the effectiveness of batch-image identification compared to standard methods. See Figure 2 for flow charts summarizing the methods employed in this study.

2.2. Data Collection

2.2.1. Field Sampling

Fishes for this study were sampled at nine sites in the Greater Toronto Area in Ontario, Canada in June 2016. Sample sites were small streams (<15 m wide) within the Sixteen Mile Creek and Rouge River watersheds (Figure 3). Each site was a 40–m section of the stream, isolated at each end using block-nets prior to sampling. Fishes were collected using three-pass backpack electrofishing (see [27] for details).

2.2.2. Fish Photography

Captured fishes were immediately placed in aerated buckets; individuals that were too large to fit into the fish viewer were identified and released without being photographed. Captured fishes were photographed in the field using a portable setup, including a folding table with adjustable legs, a digital camera (Nikon^® Coolpix L330, Tokyo, Japan) attached to a tripod with adjustable legs, and a Plexiglas^® fish viewer (a narrow fish tank with dimensions of 30 cm long × 20 cm high × 4 cm wide) to hold fishes for photography (Figure 4). All surfaces of the viewer, excluding the front panel, were made white and opaque (either by using white thermoplastic, or by covering the transparent surfaces with white duct tape) to minimize reflection and irregularities in lighting. The viewer was filled with clear water collected from the sample site, until the water level in the viewer reached approximately 5 cm from the top. After the viewer was filled, a three-sided covering of white foam board (two side panels and one lid) was affixed to the viewer using tape, to provide additional protection from light disturbance to improve image quality. A small piece of tape was affixed externally to the centre of the front panel of the viewer, to provide a focal point for the camera. A small waterproof label, with a unique identification number, was placed in a corner of the viewer, to provide an identification tag for each subsample. Flash photography was used, with the camera exposure being adjusted as necessary to compensate for light levels in the surrounding environment.

To be photographed, fishes were transferred from the aerated bucket into the viewer using a dip net. The number of individuals in an image ranged from three to 19 individuals (mean = 10), depending on the body size distribution of the fishes. Each subsample of fishes was photographed five times on average, until multiple photographs of an acceptable quality (e.g., in focus, minimal overlap of fishes) were obtained. Between photographs, the water in the viewer was gently circulated using a ruler to change the position of the fishes in the viewer; this technique was used to prevent fishes from obscuring one another, and to obtain photographs with the fishes positioned in several ways, to aid in identification.

2.2.3. Preservation and Identification

Immediately after photography was completed, each group of fishes was euthanized using a clove oil solution, with the exception of fishes too large to fit into the 240 mL sample jars, which were immediately identified and released. Three euthanasia bins were used to efficiently process fishes, as fishes took an average of approximately five minutes to be fully euthanized. Each euthanasia bin was filled with 500 mL of water, to which approximately 0.25 mL of 10% clove oil/ethanol solution was added and replenished throughout the day as required [36].

After euthanasia, each group of fishes and its identification label were placed into a 240 mL plastic jar with 10% formalin solution. After a minimum of 24 h, preserved fishes were rinsed in water for 24 h and then transferred to 75% ethanol. Ethanol-preserved fishes were identified to the species level in the laboratory. Life stage was also noted. Each identification was confirmed by at least two identifiers to ensure accuracy.

Of the 182 subsamples of fishes photographed during field sampling, 100 were chosen for further analysis based on the clarity of the images available for the subsample and the proportion of fishes visible in the images. Of the image replicates from each subsample, the highest-quality image (visually determined based on fishes being in focus, oriented perpendicular to the camera, and minimally overlapping each other) was chosen to be the basis of taxonomic identifications. Then, based on the laboratory identifications for the fishes in each batch, fishes in the images were identified to the species level. We also noted fishes known to be present in the subsample, but not visible in the highest-quality image (“completely obscured fishes”); despite efforts to prevent crowding and obscuration, the highest-quality images from some subsamples did not show all fishes present in the batch. Additionally, three small larval fishes that appeared in the survey images were unidentifiable in the laboratory and, therefore, were removed from subsequent analyses.

2.2.4. Taxonomic Expert Survey

Using the highest-quality images from the 100 selected subsamples, a series of online identification surveys were created to be distributed to fish taxonomists, to obtain image-based identifications of the fishes with which to compare the laboratory identifications. Participants invited to participate in the surveys were biologists knowledgeable in Ontario freshwater fish identification or past participants in identification courses for Ontario freshwater fishes. To ensure the surveys were not prohibitively long (i.e., to encourage completion), five images were included in each survey, for a total of 20 surveys (20 sets of five images). Surveys were numbered, and participants were asked to choose a survey to complete based on a random number generated from one to 20, in an effort to achieve an even response rate across surveys. Individual respondents were allowed to complete as many different surveys as they wanted.

The surveys first instructed participants to self-identify their level of fish identification expertise on a five-class scale, where Class 1 was the least proficient and Class 5 was the most proficient (

Table 1. Classification of relative taxonomic expertise for survey participants in this study, regarding their ability to taxonomically identify fishes in Ontario.

Expertise Class	Definition	Number of Respondents
1	Can identify all fish families in Ontario	6
2	Can identify all game fishes and common non-game fishes in Ontario	19
3	Can identify the adults of most fish species in Ontario	27
4	Can identify the juveniles and adults of most fish species in Ontario	44
5	Can identify the juveniles and adults of all fish species in Ontario	3

). Subsequently, participants were shown an image from one of the 100 subsamples and asked to identify each fish in the image to the species level or genus level if they could not determine the species (choosing from options presented in list of taxa; Table S1). Two versions of the image were shown for each subsample: a numbered image, to allow the participants to identify each fish by number; and an unnumbered image, to allow participants an unobstructed view of each fish. Five of these subsample image pairs were included in each survey. After participants completed the fish identification, they were asked to identify any difficulties encountered that prevented them from accurately identifying the fishes by indicating one or more of four standardized reasons or providing their own reason. At the end of the survey, participants were asked to suggest potential improvements to the survey and study design. The surveys were designed and conducted using Survey Monkey (see Supplementary File S1 for a sample survey).

2.3. Statistical Analysis

Unless otherwise specified, data analyses described below and their resultant figures were produced using R software ver. 4.4.1 [37] within the RStudio environment [38]. See Supplementary S2 for a complete list of packages used.

2.3.1. Time to Complete Identification

To compare the mean time to complete an identification survey (as recorded by Survey Monkey) between expertise classes, an ANOVA was conducted and, if significant, post hoc Tukey tests were run to test the significance of pairwise differences between expertise classes. Prior to conducting the tests, outlier values for completion time (>300 min; n = 7) were removed from the data, as it was assumed that such long completion times indicated that these surveys were not completed in a single continuous sitting (i.e., the survey was left open between sittings, artificially inflating completion time).

2.3.2. Accuracy of Identification

By comparing survey identifications made by respondents for each fish to the laboratory identifications, it was determined if the respondent had identified the fish correctly to the species level, had identified the fish to the genus level only (either as a different species in the correct genus, or a genus-level identification), was not confident enough to assign an identification (“unidentifiable”), or had incorrectly identified the fish (all other responses). We calculated the proportions of each of the above identification categories as

$${\displaystyle \frac{{n identifications}_i}{{n total identifications}_i}}$$

(1)

for each species i. For this and subsequent analyses, juveniles and adults were considered separately (i.e., juveniles were considered separate “species” from their adult counterparts), as the smaller size and less-distinct morphological characteristics of juveniles were hypothesized to result in lower species-level CIRs compared to adults of the same species. Mean species-level and genus-level CIRs across all taxa, weighted by number of identifications, were also calculated.

It was hypothesized that species, respondent expertise class, and fish density (i.e., the number of fishes in the image) would influence the species-level CIR for a given fish, as (1) species with more visually obvious identification features may be more easily identified from images; (2) respondents with higher levels of taxonomic expertise could be expected to have a higher likelihood of identifying a given fish correctly; and (3) more fishes in a given image would increase the possibility of the fishes crowding and obscuring one another, reducing the likelihood of correct identification. To examine the effect of these variables on CIR, a generalized linear mixed model (GLMM) was fit to the binary species-level identification data (1 for correct species-level identification, 0 for incorrect) with fixed effects for species, expertise class, and fish density (i.e., the number of fishes in the image). The model had a binomial error distribution and logit link function, with a random intercept for individual respondent (nested within expertise class) to account for multiple identifications having been made by each survey respondent. Model validation was then performed, including checking model assumptions (e.g., normality and homoscedasticity of residuals, lack of overdispersion, lack of outliers), predictor collinearity, and normality of random effects. Post hoc Tukey tests were run to test the significance of pairwise differences in CIR between different species and different expertise classes, respectively.

2.3.3. Misidentifications

Multiple analyses examining the potential contributing factors to identifications that were not correct to the species level (hereafter “misidentifications”) were also conducted. For one, the relative margin of error of misidentifications (i.e., the relative taxonomic or visual similarity of the misidentified species to the correct species) was analyzed by classifying common taxonomic misidentifications for the different species. First, the top three mistaken identifications (species or genera) were determined for each species; analyzing only the top three misidentifications minimized the influence of random mistakes or arbitrary choices on the part of the survey participants. The top three misidentifications for each species were then categorized as either a genus-level, a morphologically similar species (using photos and morphological descriptions from [39]), or an “other” misidentification. To determine the overall erroneousness of the misidentifications, the weighted mean proportions of these three categories were then calculated across all taxa. Also, to classify the difficulties with identification encountered by the survey participants, the proportion of individual identifications of each image (identifications of a given image by a given participant) with each of the four standardized reasons and custom (“other”) reasons was calculated.

Additionally, the factors contributing to participants being not sufficiently confident or unable to make an identification at all were also examined. A weighted mean proportion of “unidentifiables” (weighted by sample size), as a proportion of the total number of misidentifications, was calculated across all taxa, to investigate the relative confidence level of the survey participants in assigning identifications. Additionally, participants were unable to attempt identifications on completely obscured fishes, which were present in the subsamples but not visible in the survey images. As it was hypothesized that fish density may affect the probability of an image containing completely obscured fishes, a generalized linear model (GLM) with a Poisson distribution and log link function was fit to examine the effect of fish density on the number of completely obscured fishes in a given image.

2.3.4. Rarefaction

To evaluate the potential of using batch-image data to compute accurate species-richness estimates, rarefaction curves calculated from image-based and laboratory-based identification data were compared. Sample-based rarefaction curves were calculated for each site, using both laboratory data (including all fishes at a given site, rather than just those individuals included in the 100 survey images) and batch-image identification data. Curves were computed with 1000 randomizations. A Chao 1 richness estimate was also calculated for each site, for both datasets, using the classic form (no bias correction [40]). The biodiversity statistics software EstimateS ver. 8.0 [41] was used to calculate these curves and richness estimates. To investigate the similarity between curves and richness estimates produced by the two datasets, Welch two-sample t-tests were conducted on the mean cumulative number of species and Chao 1 richness estimates produced by curves using laboratory versus survey data.

2.3.5. Length Analysis

We investigated the accuracy of length measurements derived from batch images by comparing the lengths of individual fishes measured from the images with their lengths as measured in the laboratory. Fork length (FL) was measured to the nearest millimeter for preserved specimens in the laboratory using a measuring board, while FL was measured for fishes in the images using ImageJ [42]. If there were multiple fish of the same species in a given image, fishes from the batch images were paired to fishes measured in the laboratory on the bases of relative size and other distinguishing characteristics (e.g., colouration). FL from the images and laboratory samples was compared at the species level using a paired t-test with a two-tailed distribution (comparing length measurements at the individual level) conducted in Microsoft Excel 2016. To estimate the magnitude of differences in FL measured by each method, a weighted mean difference (with sample size as the weight) and percent difference were calculated between the lengths of the specimens measured in the laboratory versus the images. The precision of image-based FL measurements was also estimated by (1) repeating image-based length measurements three times for a subset of seven-10 individuals (for species with >10 measured individuals) or for all individuals (for species with <10 measured individuals, except hornyhead chub (Nocomis biguttatus), where two of four individuals were measured repeatedly); (2) calculating the variance between these replicates; and (3) using these individual-based variance values to calculate an overall mean of FL variance for each species. For most species, some or all individuals could not be measured because they were not adequately visible in the image (e.g., body was not positioned straight in the image or was obscured by another fish), or they had been released in the field due to their size (see Section 2.2.3). These individuals were excluded from the analysis, and the proportion of such individuals per species was quantified.

3. Results

3.1. Survey Responses

A total of 99 individuals (see Table 1 for expertise classes) completed 147 surveys for this study. Mean surveys completed per respondent were 1.48, with no respondents completing more than four surveys except for one, who completed 19 surveys. This corresponded to a mean of 71 fishes identified per respondent, with a maximum of 897.

Mean and median response times were 28.5 and 40 min, respectively (Figure 5). There was a significant difference in response times between expertise classes (ANOVA, F = 2.94, p = 0.02, n = 142), with only Class 4 having a significantly faster mean time than Class 3 (Tukey, p = 0.02). For a complete summary of respondent statistics, see Table S2.

3.2. Accuracy of Identification

3.2.1. Correct-Identification Rates by Species and Genus

The weighted mean species-level CIR across all species was 49.7%, with an additional 11.5% (total of 61.2%) of fishes being correctly identified to the genus level. Species-level CIRs varied significantly by species and life stage (

Table 2. Analysis of deviance table showing the results of Wald II chi-squared tests, based on the results of a generalized linear mixed model examining the effect of species, expertise class, and fish density on the probability of correct species-level identification. Significant p-values (α = 0.05) are marked with an asterisk (*).

Predictor	Chi-Squared	df	p-Value
Species	764.144	24	<0.001 *
Expertise Class	39.897	4	<0.001 *
Fish Density	0.455	1	0.500

), ranging from 5.4% for juvenile creek chub (Semotilus atromaculatus) to 87.5% for largemouth bass (Micropterus salmoides) (Figure 6). Juveniles and species with cryptic identification features (e.g., central stoneroller (Campostoma anomalum), cartilaginous ridge on lower jaw; mottled sculpin (Cottus bairdi), presence of palatine teeth [39]) generally had significantly lower species-level CIRs compared to larger and more morphologically distinct species (e.g., rainbow trout (Oncorhynchus mykiss), white sucker (Catostomus commersonii); Table S3). Genus-level CIRs did not follow the same rank order as species-level CIRs, as some species with cryptic identification features (e.g., bluntnose minnow (Pimephales notatus), variation in scale size [39]; mottled sculpin) were correctly identified to the genus level (i.e., species-level identification or genus-only) at substantially higher rates than to the species level (Figure 6). Cases where species- and genus-level CIRs were equal for a given species (e.g., central stoneroller) occurred due to a lack of a genus-level identification option for that species in the survey because there is only one species in the genus in Ontario.

3.2.2. Factors Affecting Correct Species-Level Identification

Results of the GLMM predicting the probability of correct identification to the species level demonstrated significant effects of species and expertise class on probability of correct identification, while fish density did not have a significant effect (Table 2; Table S4). Many pairwise differences in effect size were significant between species, particularly for species with low CIRs, high CIRs, and those with large numbers of identifications (Table S3). Pairwise differences between expertise classes were often significant, with the exception of those between neighbouring classes and between Class 1 and other classes (Table S5). CIR increased continuously from Class 2 to Class 5, but Class 1 survey participants achieved a mean CIR significantly higher than Class 2 or 3 (Figure 7).

3.2.3. Misidentifications

Across all species, the most common misidentification was “unidentifiable,” which comprised 46% of all misidentifications and was particularly common for species with low CIRs (Figure 5). Within the top three misidentifications (excluding unidentifiable) for each species, the most common misidentifications were genus-level identifications (51%), followed by “other” misidentifications (31%), and species morphologically similar to the target species (18%) (Table S6).

GLM results showed a positive and significant relationship between fish density and completely obscured fishes (

Table 3. Effect of fish density (subsample size) on the number of completely obscured fishes in an image. Estimated regression parameters, standard errors, z values and p-values for the Poisson generalized linear model. Significant p-values (α = 0.05) are marked with an asterisk (*).

Predictor	Estimate	Std. Error	z Value	p-Value
Intercept	−2.751	0.502	−5.480	<0.001 *
Subsample Size	0.182	0.039	4.679	<0.001 *

). There were no completely obscured fishes in subsamples with six or fewer individuals, and the mean number of completely obscured fishes increased substantially at sample sizes of greater than 10 (Figure S1). The three most common identification difficulties self-reported by the survey participants were low image resolution (51%), poor fish orientation (51%), and obscured identification features (44%). Poor image quality (22%) (for reasons other than resolution, such as lighting) and self-described “other” problems (11%) occurred less frequently (Figure S2).

3.3. Rarefaction

Rarefaction curves (Figure S3) for each site produced by batch image data had slightly lower (t (192) = −2.15, p = 0.03) mean cumulative species estimates (M = 7.36, SD = 2.17) than those produced using laboratory data (M = 7.97, SD = 1.83). However, Chao 1 richness estimates were not significantly different between the two groups (t (16) = −0.54, p = 0.60).

3.4. Length Analysis

There were generally small differences in FLs measured from batch images and laboratory specimens (Figure S4, Table S7), with specimens being measured a weighted mean of 3.79 mm larger in the laboratory than in the images (representing a mean difference of approximately 6%). Differences in measurements of FL based on batch images were statistically significant for four of the 13 species, as determined by paired t-tests (p < 0.05) (Table S7). Across species, a mean of approximately 52% of individuals were not measured, almost exclusively due to inadequate visibility in the image (Figure S5).

4. Discussion

In this study, batch-image identification as a method for taxonomic identification of freshwater fishes was found to be less accurate than conventional manual identification. However, data from batch images produced species-richness and length estimates that were comparable to those obtained from laboratory identification. Furthermore, methodological improvements could substantially improve the accuracy of batch-image identification.

4.1. Identification Accuracy

The overall mean species-level CIR from batch-image identification of approximately 50% was substantially lower than the high-accuracy [9] method of conventional manual identification by a fish taxonomy expert (an estimated accuracy of 90%; N. Mandrak, pers. obs.). However, CIR varied significantly and substantially between species. Large, morphologically distinct species had relatively high CIRs (e.g., adult rainbow trout: 80%; adult white sucker: 68%). Mid-sized, less morphologically distinct species had more moderate CIRs (e.g., rainbow darter (Etheostoma caeruleum): 55%; blacknose dace (Rhinichthys atratulus): 55%). Juveniles and morphologically indistinct species with cryptic identification features had low CIRs (e.g., central stoneroller, 10%; juvenile white sucker, 10%). Although some of this variability may be explained by large heterogeneity in sample size (e.g., stonecat (Noturus flavus): n = 8 identifications, versus rainbow darter: n = 1114 identifications), the significant variability in species-level CIRs is likely a result of differences in size and morphology between species. Thus, batch-image identification appears to be effective as a method of species-level taxonomic identification for large and/or morphologically distinct species, but may require methodological improvements to achieve high species-level CIRs for smaller and less morphologically distinct taxa. However, small and morphologically cryptic species are often difficult to identify even when conducting conventional field-based identifications [3], so this issue is not unique to batch-image identification.

Although species-level CIRs were low for some species, genus-level CIRs were generally higher than those at the species level. The overall mean genus-level CIR (61.2%, when combining correct species- and genus-level identifications) was substantially higher than the overall species-level CIR and, in particular, species with cryptic identification features (e.g., bluntnose minnow, mottled sculpin) were identified at much higher accuracy to the genus level than they were to the species level. Juveniles were the only fishes (of those that had a genus-level identification option available in the survey) that showed little difference in CIR between the species and genus levels. Taxonomic identification to the genus level or higher is often adequate for fish surveys, depending on study goals [43]. For example, Vanderklift et al. [44] found that genus richness was reflective of species richness when surveying marine fish assemblages. Therefore, batch-image identification may be especially suitable as an alternative to conventional identification methods for studies requiring taxonomic identification only to the genus level or above.

CIR also varied by taxonomist expertise, with our results showing a general trend of increasing species-level CIRs with increasing identifier expertise class. Survey respondents in the highest taxonomic expertise classes were able to identify fishes from batch images at elevated rates (59%: Class 4; 83%: Class 5) compared to the average CIR across expertise levels (~50%). However, Class 1 respondents showed significantly more accurate identifications than Class 2 or 3 respondents, and Tukey tests demonstrated that Class 1 CIR was not significantly lower than CIR for Classes 4 and 5. The higher CIR of Class 1 respondents over those from Class 2 and Class 3 is likely an artifact of some Class 1 respondents either underestimating, or Class 2 and 3 respondents overestimating, their taxonomic identification expertise, or some Class 1 respondents misinterpreting the expertise scale and mistaking Class 1 as the highest expertise category. Additionally, a relatively low number of identifications made by Class 1 respondents may have decreased the statistical power of the Tukey tests, leading to non-significant differences in CIR compared to other categories (including Class 5, which itself had the lowest number of identifications made). Therefore, the overall trend of increasing CIR with increasing expertise is likely valid, and expert taxonomists can achieve higher CIRs using batch-image identification.

Batch-image identification produced population and community data comparable to laboratory methods. Length measurements based on the batch images did not substantially differ from laboratory-based length measurements, with a low magnitude of discrepancy (<4 mm) from the laboratory measurements even when statistically significant differences were present. The mean margin of error of ~6% was only slightly higher than that of some past studies estimating fish length from images (e.g., 3% [45]) and could be explained by the effects of factors such as viewing angle [46] and image distortion (due to refraction [33]) on length measurement, and the challenge of obtaining clear images of multiple live fishes due to fish movement (most past studies involving fish photography photographed fishes either dead or immobilized [34]). Furthermore, batch-image identification consistently slightly underestimated FL, which could allow for the development of a correction factor that could be applied to batch-image-derived FL measurements to improve their accuracy. Although some fishes were unable to be measured, often due to issues relating to their visibility within the batch images, the similarity of our FL measurements to laboratory measurements shows that accurate estimates of fish length can be obtained from batch-image identification. However, studies using batch images to obtain fish length measurements should ensure adequate sample sizes to account for a portion of the fishes being unmeasurable.

In addition to population demographic data such as fish lengths, our study also found that batch-image identification data can produce reliable estimates of species richness. Species-richness estimates from rarefaction curves were only slightly underestimated compared to laboratory identification data, and species-richness estimates generated from batch-image identification data did not differ significantly from those produced using our complete laboratory sample data. This indicates that accurate approximations of biodiversity can be obtained using batch-image identification, expanding the scope of its potential utility.

Our results for the most common misidentifications show that survey respondents did not err far from the correct species-level identification for a given fish when they were confident enough to provide an identification (as opposed to marking the fish as “unidentifiable”). Most commonly, our survey participants “misidentified” a fish by providing a genus-level identification (e.g., identifying a blacknose dace as “Rhinichthys spp.”). For multiple species with cryptic identification features, this likely occurred because survey participants were not confident in identifying the fish to the species level, as evidenced by lower species-level than genus-level CIRs for cryptic species. Participants also often arrived at an incorrect identification by mistaking the fish for a morphologically similar species. This indicates that small improvements to the batch-image identification methodology could produce considerable improvements in species-level CIR if they lead to slight improvements in identifiers’ ability to discern details of fish identification features.

4.2. Processing and Identification Effort Compared to Conventional Sampling

Based on our experience in this study, for a river site with average diversity of five–10 species and capture of about 100 individuals total, it takes about 0.5–1 h to conventionally process (euthanize, preserve) fishes on site and 1–2 h to process (rinse, identify, catalogue/discard) in the lab (1.5–3 h/site; mean = 2.25 h). Our digital method took about 0.5 h in the field and a mean of approximately 1 h to identify ~100 fishes on average from 10 images (i.e., two surveys; total of 1.5 h/site). Thus, batch-image identification represents a more time-efficient alternative to conventional identification methods.

4.3. Limitations

There were several limitations to our study. Despite our efforts to minimize fish obscuration, highest-quality images for some batches contained some fishes that were obscured by other fishes and, therefore, could not be identified. However, this allowed us to determine the optimal number of fishes in an image to avoid overlapping individuals (see below). The digital camera used in this study was a point-and-shoot camera, with 20.2-MP resolution and no manual shutter-speed control, which limited image resolution and quality. The white sides and background used to control light were damaged during transit and resulted in varying image quality between sites. Using the ruler to move fishes did reduce water clarity and, hence, image quality. We identify potential solutions to these limitations below.

4.4. Methodological Improvements

There are numerous potential ways in which the methodology of batch-image identification and, consequently, its accuracy and feasibility, could be improved. Potential improvements can broadly be classified as improvements to the procedures used to collect the batch images and improvements to the identification of fishes from the images.

4.4.1. Image Collection

Although our analyses did not find a significant relationship between batch size and CIR, our results did indicate a positive relationship between the number of fishes in a subsample and the number of fishes not visible in a given batch image. Additionally, this phenomenon was supported by observations made during the image-collection component of this study; as more fishes were put into the viewer, it was more difficult to obtain an image in which all fishes were visible (C. Pratt, pers. obs.). As completely obscured fishes are impossible to identify through batch-image identification, the accuracy of batch-image identification could, therefore, be improved by including fewer fishes per image subsample. Limiting the number of fishes in each subsample to a critical value would reduce the likelihood of fish obscuration. In our dataset, there were no completely obscured fishes at batch sizes of six or fewer individuals and minimal completely obscured fishes below an inflection point of 10 fishes. Therefore, we would recommend these values as conservative and maximal batch sizes, respectively.

In addition to changes to the procedures for batch-image identification, improvements to the equipment used could result in better images and, therefore, improved accuracy. Using a camera with higher resolution and the ability to control shutter speed could improve image detail (by making identification features easier to see) and image quality (by reducing blurring caused by fish movement), respectively. Alternatively, a high-quality smartphone camera with a macro lens could be used to improve image quality, which would allow for the added benefit of real-time image upload to cloud storage to improve the efficiency of file storage and minimize the possibility of data loss. Secondly, the use of a ruler to prevent schooling and alter fish orientation by moving the water in the viewer had the negative side effect of causing turbulence in the water, which often led to blurred images, and was only temporarily effective, often allowing fishes to re-school before an image could be taken. A thinner fish viewer could present a potential alternative to the ruler method, reducing schooling and poor fish orientation. Reducing the width of the fish viewer would inhibit the ability of fishes to school together and obscure one another and also prevent them from orienting themselves perpendicular to the front wall of the viewer, which would result in more consistent visibility of their identification features. However, this would present a trade-off of limiting the potential size range of fishes that could fit in the viewer, so viewer size would need to be tailored to the goals of a given study. Lastly, the foam boards used in this study to control light were liable to shift and become damaged while being transported and handled during sampling. This reduced their effectiveness at controlling light levels in the fish viewer and led to more inconsistent image quality across the subsamples, inhibiting the ability of the survey participants to identify the fishes in the images; poor image quality was reported as an identification difficulty in 22% of survey images. In future implementations of batch-image identification, a more consistent and durable light-control method (e.g., hard plastic sheets attached via a hinge) would improve light control in the images taken and, thereby, improve the CIR.

4.4.2. Fish Identification

There are multiple potential improvements over the taxonomic identification procedures used in this study that could improve the accuracy of batch-image identification. For one, the trend of higher CIRs with increasing taxonomic expertise indicates that accurate image-based fish identification requires a relatively high minimum level of taxonomic expertise. Therefore, inclusion of only high-expertise identifiers (Class 4, Class 5) in future implementations of batch-image identification would improve its accuracy. Although this would not solve the challenge of finding taxonomic experts with high expertise [9], batch-image identification would remain advantageous over manual identification in this case, as the expert need not be present in the field since images could be examined remotely.

Additional improvements to the effectiveness of batch-image identification by taxonomists could be achieved by altering the characteristics of the images and related information provided to identifiers. In this study, survey participants were not provided with contextual information about the fish samples, such as where and when they were collected. This information was withheld to avoid biasing the identifications of the survey participants. However, in a practical implementation of batch-image identification, this information would be known to the identifiers and would likely aid in their identifications. Similarly, allowing identifiers to examine multiple images from a given subsample, or even a short video instead of images, could improve their identification accuracy and, therefore, the CIR. This approach would have to balance accuracy with time invested as examining multiple photos and/or videos of each subsample would require additional time to be spent on each subsample by the identifier.

Given the limitations of human-based identification in terms of both accuracy and resource constraints, batch-image identification may be improved by being paired with automated taxonomic identification. Batch-image identification provides a high level of consistency in imagery, a known species pool for which to develop taxonomic algorithms, and images of multiple fishes with minimal overlap, making it potentially amenable for integration with AI-based identification methods. However, batch imaging will always result in some degree of inconsistency in image quality (due to images being taken in the field), images will contain an array of taxa and life stages, and images will always contain multiple individuals. Therefore, due to the limited ability of current AI-based identification approaches to handle inconsistency in image quality, variation in fish morphology, and images with multiple fishes [26], further research and advancements may be required before batch-image identification and AI-based methods can be effectively integrated.

Although there are multiple ways in which batch-image identification could be modified and improved as an identification technique, the inherent limitations of visual taxonomic identification (i.e., lack of ability to easily differentiate visually similar species, especially early life stages [47]) can only be surmounted through the use of non-visual identification methods. Pairing batch-image identification with molecular identification, such as eDNA metabarcoding or targeted quantitative PCR, could result in a complementary sampling program with higher reliability and identification accuracy than either technique alone, as has been found when pairing eDNA with other visual survey techniques e.g., [48,49,50]. Moreover, batch-image identification could be particularly amenable to pairing with eDNA sampling. For one, eDNA identification could surmount one of the main limitations of batch-image identification—the difficulty of identifying juvenile fishes and fish species with cryptic identification features. Secondly, batch imaging of fishes already involves the collection of small volumes of water paired with each batch of fishes; after imaging, the water from each subsample could be filtered for eDNA analysis to aid in confirming taxonomic identification and also identifying non-captured species in the waterbody. Alternatively, because analyzing water from each subsample could be resource-intensive, eDNA samples could simply be obtained separately from the same waterbody as the captured fishes. Overall, in addition to its potential use as a stand-alone identification method, batch-image identification could also be a complementary addition to the existing array of taxonomic identification techniques for freshwater fishes.

5. Conclusions

Batch-image identification represents a novel, efficient, easily employable alternative to standard freshwater fish identification methods, with the potential to produce accurate identifications and ecological data with a high level of accuracy. Batch-image identification offers solutions to the shortcomings of standard manual identification methods by minimizing fish mortality from handling and physical voucher collection, providing digital vouchers, and alleviating the need to have a fish identification expert present in the field. It also has advantages compared to other alternative methods to manual identification; for example, it is more efficient than the predominant image-based identification methodology of imaging fishes one individual at a time and provides more ecological information than eDNA methods. Although batch-image identification resulted in lower overall identification accuracy than standard manual identification, our study provided a successful proof-of-concept of batch-image identification and also yielded quantitative and qualitative observations supporting further improvement of this methodology. There is also the opportunity to combine batch-image identification with other methods currently used in fish identification, such as AI-based identification and eDNA sampling, in a complementary manner whereby batch-image identification and these other methods could mutually address their respective shortcomings. Given the potential utility of batch-image identification, future studies are recommended to further the development of this method.