Northeastern Naturalist 192 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 22001144 NORTHEASTERN NATURALIST 2V1(o2l). :2119,2 N–2o0. 92 Application of Mobile Dual-frequency Identification Sonar (DIDSON) to Fish in Estuarine Habitats Kenneth W. Able1,*, Thomas M. Grothues1, Jenna L. Rackovan1, and Frances E. Buderman2 Abstract - Dual-frequency identification sonar (DIDSON) offers important advantages over other sampling tools for observing pelagic and benthic fishes in situ. Because it relies on sound, DIDSON can detect fish in a non-destructive and non-intrusive manner. In our unique application, the equipment’s small size and low power requirements allow deployment from a kayak for increased maneuverability in complex habitats. Characteristics that typify echograms of different fishes can be extracted using multivariate ordination techniques, such as principal components analysis (PCA), with in situ groundtruthing. Here we present reference images, techniques, and human-observer–error estimates from DIDSON application. Together, these approaches enhance our ability to sample fishes and even observe certain behaviors in complex, turbid environments during a full diel cycle. Introduction Our basic lack of knowledge of natural history limits our understanding of the fauna in many environments (Arnold 2003, Cotterill and Foissner 2010, Dayton and Sala 2001), particularly in aquatic habitats and especially so in turbid waters such as estuaries. Conventional sampling techniques are inappropriate for nektonic fishes in complex, typically turbid estuaries (Brehmer et al. 2003). Under-pier habitats cannot be sampled effectively with towed nets because pilings offer refuge for fish (Stoecker et al. 1992), and snags prevent the use of either towed or set (gill) nets, especially in areas of flow or vessel wake. Larger fish easily avoid cast or pop nets that are suitable for smaller schooling fishes. Hook-and-line fishing can potentially determine the species occupying these habitats, but it is non-quantitative and biased to certain species and to particular times of the day and locations, when and where fish are feeding. Further, the expected statistically rare occurrence of large fishes in any habitat requires a wide spatial and temporal sampling effort, which is rarely spatially discrete. Visual census (by divers or video; e.g., Bortone et al. 2000) is potentially rapid and discreet, but is limited by turbidity, lack of light at night, and safety issues in many estuaries, especially those in urban settings. Further, the response of pelagic fish presents some serious challenges to further study because of their great mobility and clumped distribution due to social behavior and schooling. Dual-frequency identification sonar (DIDSON) offers many capabilities to overcome these difficulties. 1Marine Field Station, Institute of Marine and Coastal Sciences, Rutgers University, 132 Great Bay Boulevard, Tuckerton, NJ 08087-2004. 2Department of Fish, Wildlife, and Conservation Biology, Colorado State University, 1474 Campus Delivery, Fort Collins, CO 80523. *Corresponding author - able@marine.rutgers.edu. Manuscript Editor: Thomas Maier Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 193 In prior studies, DIDSON has been used to characterize habitat (Tiffan et al. 2004), count and measure fish (Becker et al. 2011a, Boswell et al. 2008, Holmes et al. 2006), study fish and jellyfish behavior (Crossman et al. 2011, Han and Uye 2009, Kimball et al. 2010, Pavlov et al. 2009, Xie et al. 2008), and estimate escapement (Holmes et al. 2006). All of these studies were based on static deployment of DIDSON, with the exception of Handegard and Williams (2008) and Able et al. (2013). Further, many studies focused on the development of automation in quantification, which is usually applicable in situations of static or homogenous backgrounds. Our application of DIDSON, deployed from a moving kayak, provided the opportunity to survey multiple habitats and their associated fauna over short time-spans in complex habitats. The purpose of this paper is to present some of the advantages and disadvantages of DIDSON in a mobile application. More specifically, we identify fish features that are useful for target classification and provide a preliminary assessment of the precision of target classification between observers when the use of automated routines is counterindicated by moving and varying ba ckgrounds. Materials and Methods Study area A portion of the study was carried out in the boat basin of the Rutgers University Marine Field Station (RUMFS) in Great Bay near Tuckerton in southern New Jersey. Other observations occurred in New York Harbor along sections of the eastern shore of the Lower Hudson River along Manhattan. Primary surveyed habitat at this latter site included overwater piers and uncovered pile fields from relict or collapsed piers interspersed with open-water areas. See Able et al. (1998, 2010) for further description of these study sites. Acoustic techniques We used DIDSON (Sound Metrics Corp., Seattle, WA) to image fishes. As used with 96 beams in the high-frequency (1.8 MHz) mode at the applied range of 1.25–10 m out from the lens, the width resolution varied between 6.5 mm and 52.1 mm per pixel, respectively. The pixel-height resolution depended on the window length, and at the applied settings of 2.5-m to 10-m window lengths, the height resolution varied between 4.8 mm and 19.5 mm per pixel, respectively. Thus, a fish 250 mm in length that was oriented in profile across the screen and 2 m from the lens was represented by an image 24 acoustic pixels long, but only 4 pixels long if it were 10 m away from the lens. At this resolution, objects as small as 40 mm could be readily identified as fishes in the very near field, but would be represented by only 1 pixel in the extreme downrange. Classification of fish and targets resembling fish required additional information or closer inspection. At that minimum size, individual fish were difficult to discern and could not be measured, but schools or aggregations were apparent. Sampling was conducted at a rate of 5–10 frames per second (fps), dependent on range, because range setting affected the host software’s processing speed. A moderate frame rate, typically 7 fps, helped discern moving fishes from a moving background (actually a static background moving relative to a moving viewer). Fish movement can be diagnostic, helping to break fish outlines from their background. Northeastern Naturalist 194 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 DIDSON images can even detect individual fins in larger fish, which generally have low reflectance, but which are valuable to identification (Brown et al. 2007). Existing DIDSON host software (Sound Metrics Corp. version 5.14) supports several tools to help measure and count detected objects (DIDSON 2006). In this study, the DIDSON was mounted with a hinge directly under and behind the bow of a sit-on-top kayak (4 m) for easy access under piers and around pier pilings. Based on earlier trials, we set the tilt at about 23° for an optimum viewing range in front of the kayak. This arrangement allowed us to view the structural components of the habitat bottom, those that extended in the water column (e.g., pier pilings), and fishes on the bottom and in the water column at both study sites. We carried a splashproof laptop computer in the kayak cockpit, which allowed real-time viewing so that the paddler could adjust focus and direction for closer inspection of potential targets. The paddler used a small red headlamp for navigation during nighttime sampling and under the darkest piers to minimize the potential effect on the fishes. A nearby motor skiff, kept outside of the immediate study area, provided logistical support and safety in the Hudson River study site. Because GPS did not function under the piers, we noted the position of the kayak using vocal annotation supported by DIDSON host software to link it with the video. To map fish position, we correlated time stamps from DIDSON recordings to navigation recordings. Although waters at the study sites were often calm, especially under large piers or in relict pile fields, episodic ship wake produced serious pitch and roll, causing smearing artifacts on the echograms. We note here that use of the term echogram for this video-like dynamic image differs from its use regarding the standard static echogram where the x-axis is time. The paddler had to stabilize the kayak during these short but intense disturbances, briefly interrupting the forward progression along the transect. Such disturbances, however, were always either preceded or followed by calm periods, so these events never prevented us from collecting steady echogram sequences useful for measuring and counting fish. Groundtruthing and image analysis In an attempt to determine the congruence between observations made with DIDSON and more traditional sampling devices, we compared several types of sampling gear during the summers of 2007 and 2008 at RUMFS. We sampled 73 DIDSON transects during the day and 41 at night ranging from 30 sec to 5 min in duration in order to capture both large and small pelagic fishes. Samples with comparison gear included a pop net (n = 16 deployments; Hagan and Able 2003), experimental traps (n = 30; Able et al. 1998), cast nets (n = 16; Johnston and Sheaves 2008), and gill nets (n = 3; Able et al. 2009) during day and night. We used visual observations (n = 60) from the surface (lying on dock next to the kayak while DIDSON was recording) and while snorkeling. Pop nets, cast nets, and gill nets were designed to capture fishes in the water column and the experimental traps were designed to capture benthic fishes. Sampling at night included the use of lights to attract and retain fishes, per Hagan and Able (2008). We identified and measured all of the fishes captured with groundtruth sampling gears. In other instances, we tethered Centropristis striata (Black Sea Bass) and Tautoga onitis (Tautog) to Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 195 monofilament line and deployed them in front of the DIDSON sensor to record their images as reference images for later identification of free-ranging animals. We compared all of these images to DIDSON deployments from a kayak (n = 114). These groundtruth samples targeted individual fishes and schools for simultaneous DIDSON imaging and capture with various gears, and we analyzed them to determine the accuracy of our identifications and to train observers. The task was simplified because of the broad overlap between the fish faunas of the two groundtruthing areas (Able and Fahay 2010). Features of the fish in both study areas (their position in water column, length, spacing, and orientation) provided metrics for identifying echogram targets. We stored DIDSON files with time stamps and metadata in native format and reviewed them in the laboratory as movies under direct playback control by the reviewer. Playback controls included frame-advance rate and direction, magnification (zoom), and graphic user interface tools such as super-position of a grid to aid counts, and point-drag paths to measure fish lengths. Two or three independent reviewers viewed the files, then scored and compared their results. Reviewers recorded each fish-presence event (either school or individual fish). We determined abundances manually or, for large schools, by making an estimate using the average number of fish per grid square based on counts of 3 squares (using the superimposed grid) and multiplying that by the number of squares with fish in them. We also took a range of measurements for length and body depth of the fish. We calculated both fish depth in the water column and bottom depth using the following formula based on the sonar tilt and hypotenuse (tar get distance): cos (radians (90 - sonar tilt)) * hypotenuse If the event was a fish school, then reviewers also measured the nearest-neighbor distance for several individuals chosen at random from the school. Reviewers also categorized school organization on a ranked basis of 1–4 from highly organized (parallel swimming, reaction to nearest neighbor as in a school) to random milling (as in an aggregation; see Pitcher 1983) as a potential metric for identification of taxon or category. Each reviewer of the DIDSON files used the matched audio file and written notes, as well as underwater landmarks visible in the echogram, to reference the position of the fish relative to pilings or other habitat structures. The reviewer also scored the amount of debris found on the bottom using a range from 1–4 (low to high) as well as qualitatively categorizing small and large debris as a measure of structure. Because we wanted to evaluate the potential effect that observer experience could have on evaluating DIDSON imagery, we compared results of individual experienced reviewers against those of an inexperienced reviewer. We assessed both fish abundance and number of events separately because each is processed visually by different learned skills. Ordinary least-squares regression of counts (or events) of any one reviewer against those of another for the same file should have resulted in a slope (regression coefficient) of 1 if the files were scored the same by both reviewers. We examined the dispersion of residuals and coefficient Northeastern Naturalist 196 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 of determination (r2) even when the slope equalled 1 to determine if the slope was caused by uniformly similar scores (low error and residual dispersion) from both reviewers or if errors were large but compensating (unbiased). In a second analysis, we examined the change in error variance as a function of the number of reviewers (2 or 3). Because individuals in larger schools of fish are potentially more difficult to enumerate than those in smaller schools, we standardized error variance to the coefficient-of-variation (CV). If the use of multiple independent reviewers was beneficial regardless of experience, then CV should have decreased as reviewers increased. We chose the files using ranks assigned by a random-number generator in MS Excel and included those with both relatively sparse and abundant targets on various backgrounds. We used the data captured during reviewer-training efforts at RUMFS, when fish were observed by eye and captured in front of the instrument during DIDSON recording (i.e., known fish), in a principal components analysis (PCA) to quantify the importance and confidence in differentiating images into fish taxa/categories. PCA is a multivariate iterative regression (taxa against image variables and vice versa) that reduces variation in the composition of samples to the most important latent gradients (eigenaxes). Graphing the regression coefficients of samples of major eigenaxes against each other produces an ordination diagram that depicts both the gradient in character values (such as image depth to length) relative to different taxa and also the relative contribution of the different characters in defining that order. Further, the percent variation explained by each of the eigenaxes is quantified as an eigenvalue (ter Braak and Smilauer 2002). Because the data taken from each image during review encompassed much of the information that was available to a reviewer who had not necessarily seen the actual fish, the PCA ordination provided a way of communicating to readers the relative value of these measures of different characteristics to the viewer classification-process. Although it is possible to write PCA algorithms into software code that help classify targets in near-real time (inclusive of, but not limited to sonar imaging; see Soares et al. 2001, Turhan-Sayan 2005), we did not attempt that nor did we apply PCA a posteriori to unknown targets. Results Image analysis DIDSON imaging from the kayak during initial groundtruthing at RUMFS provided mobility when maneuvering around surface structures and clear imagery of the structural components of habitats. During these deployments, bulkheads, pier pilings, and remnants of crab pots (rebar frames) were clearly visible on the bottom and in the water column. We investigated other structures in New York Harbor based on the same mobility. These objects included miscellaneous dense (Fig. 1a) and light (Fig. 1b) debris, and an I-beam (Fig. 1c) on the bottom. A potentially confounding factor in interpretation of DIDSON images was the frequent release of bubbles from the organically rich sediment in these urban habitats (Fig. 1d). These were easily resolved from fish in video format, but were confusing in static images. Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 197 For the fish fauna, reviewers documented as many as 5 events (different occurrences of fishes as individuals or schools or both) on single transects. Reviewers recorded an event when they detected objects with a discontinuous occurrence and orientation to different structures within the transect. However, given the mobility of large fish and the relatively narrow viewing area of the DIDSON, reviewers used their judgement to determine the independence of these events. The difference in judgment among experienced and inexperienced reviewers accounted for much of the experimental error in the number of events counted, with the coefficients of variations higher for the abundance counts than for event counts (Fig. 2). Regardless of the type of response measure, the addition of an inexperienced reviewer resulted in a higher CV than we calculated when there were two experienced Figure 1. DIDSON images of various structural components on the bottom in the Hudson River study area. The numbers along the side of each image indicate the distance (m) from the sensor. The image includes: a) debris (high score of 4) in a piling field near Pier 57, b) light debris at the edge of a bridge on the New Jersey side of the Hudson River (low score of 2), c) no debris by a piling under Pier 40 (score of 1) except for I-beam, and d) vertical bubbles (shown by white ovals) rising from the sediment surrounded by a school of small fish. Northeastern Naturalist 198 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 reviewers (Fig 2). Thus, additional review by an inexperienced reviewer did not increase confidence in the solution obtained by independent, experienced reviewers. In the regression analysis, the slopes were near unity (c = 1.2) when comparing fish counts among two experienced reviewers, and less so (c = 1.6) when comparing event counts among experienced reviewers. The slopes were similar if one of the reviewers was experienced and one was inexperienced (c = 1.2 and 1.3 for count and event count, respectively). However, the coefficient of determination was much stronger for experienced reviewers (r2 = 0.65 and 0.68, vs. 0.50 and 0.30, for count and event count, respectively). This result indicated that the error variance was much greater when using inexperienced reviewers, but that the error was unbiased; reviewers counted both too many and too few fish (or events) in equal proportion. The splitting or lumping of events did not change the overall estimate of abundance. For example, one reviewer might have counted 3 individual fish as 2 events (e.g., a singleton and a pair), while another reviewer counted 3 fish as 3 events. Both reviewers arrived at the same fish count (i.e., 3); however, this could affect the average count per event, which was not scored in our study. Another source of variation in abundance came from the estimation of total number from subsets in large schools. Identification of fish taxa and categories We compared the results of fish identifications made using DIDSON imagery collected during groundtruthing at RUMFS with the results from samples collected using a variety of gears including pop nets, cast nets, and gill nets for pelagic fishes, Figure 2. Average (bar) and standard error (error bars) of fish-response measures (abundance and events) as counted by 2 or 3 reviewers for a random selection of DIDSON files. The average coefficient of variation for each case is also shown. In all test cases, two reviewers were experienced and another was considered an inexperi enced reviewer. Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 199 and experimental traps for benthic fishes during the day and night (Table 1). Of the 15 species/taxa collected across all groundtruthing gears in the RUMFS boat basin, reviewers of the congruent DIDSON files identified only 5 fish species, including Brevoortia tyrannus (Atlantic Menhaden), Black Sea Bass, Menidia menidia (Atlantic Silverside), Tautog, and Pomatomus saltatrix (Bluefish) (Table 1). It is perhaps not surprising that reviewers of DIDSON imagery did not identify several less-common species (e.g., Alosa mediocris [Hickory Shad] and additional Tautog individuals). However, several abundant species were also not identified—Anchoa mitchilli (Bay Anchovy), Alosa pseudoharengus (Alewife), Fundulus heteroclitus (Mummichog), Sphyreana borealis (Northern Sennet), and Tautogalabrus adspersus (Cunner). In addition, Bay Anchovy were of the same size and mixed with aggregations of the more abundant Atlantic Silverside, as identified by cast-net samples in front of the DIDSON, but individuals of these species could not be differentiated in DIDSON files. However, we identified a small shark with the DIDSON during groundtruthing at RUMFS (likely Mustelus canis (Mitchill) [Smooth Dogfish] , based on the presence of a heterocercal caudal fin and the subcarangiform swimming motion), which was not collected in groundtruthing nets. Several species collected in groundtruth sampling were too small to have been detected with DIDSON (e.g., Gobiosoma sp. [goby]) or cryptic on the bottom or buried (e.g., Pseudopleuronectes americanus [Winter Flounder]). Thus, the ability of DIDSON to identify the diversity of estuarine fishes is limited to a few distinctive forms and, in this application, it was easier to detect and identify lar ger individuals. When these forms were subdivided by some aspects of behavior (benthic, pelagic) and size, using a cutoff for maximum size of mature Atlantic Silverside and Bay Table 1. Fish-species composition and relative abundance by groundtruthing and from DIDSON images based on sampling in the Rutgers University Marine Field Station boat-basin and nearby areas in southern New Jersey. x indicates species identified but not counted. Groundtruth method Pop Cast Gill Species/Taxon net net Trap net Visual DIDSON Anchoa mitchilli Valenciennes (Bay Anchovy) 8 1 0 0 0 Anchoa sp. 1 0 0 0 0 Alosa pseudoharengus Wilson (Alewife) 1 11 0 0 0 Alosa mediocris Mitchill (Hickory Shad) 1 0 0 0 0 Brevoortia tyrannus Latrobe (Atlantic Menhaden) 1 0 0 0 0 x Centropristis striata L. (Black Sea Bass) 0 0 3 0 x x Fundulus heteroclitus L. (Mummichog) 0 2 24 0 0 Gobiosoma sp. 1 0 2 0 0 Menidia menidia L. (Atlantic Silverside) 232 2 2 0 x x Pomatomus saltatrix L. (Bluefish) 8 5 0 3 0 x Pseudopleuronectes americanus Walbaum 1 0 1 0 0 (Winter Flounder) Sphyraena borealis DeKay (Northern Sennet) 14 0 0 0 0 Tautoga onitis L. (Tautog) 0 0 1 0 x x Tautogolabrus adspersus Walbaum (Cunner) 0 0 16 0 2 Unidentified Clupeid 2 1 0 0 0 Northeastern Naturalist 200 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 Anchovy (<250 mm), we were able to sort them into reasonably discrete categories that could be reliably replicated (Table 2). We made our distinction between schools and aggregations (Fig. 3) based on the organization of individuals relative to each other. Schools included objects of similar size that were tightly organized with small distances between individuals. Aggregations were often comprised of slightly dissimilarly sized objects with multiple orientations of individuals. We classified schools and aggregations as large if they contained >50 individuals, otherwise we considered them to be small. We determined that these categories would be useful for gauging responses of fish in estuaries and for use in other complex habitat assessments in future studies. Although the details necessary for identification with DIDSON were less distinguishable for smaller fish, behaviors observed in the RUMFS boat-basin added to our ability to identify categories or even species. We compared variation in school organization, size distribution, and individual fish spacing in echograms of schools of small fish to composition of individuals in net captures during groundtruth efforts and visual observations at RUMFS (Table 1). Resolution of these known fish by PCA was moderate; the primary eigenaxis explained 29% of the variation, the second eigenaxis explained an additional 15%, and the third and fourth explained 14% and 12%, respectively (Fig. 4, Table 3). Thus, a 2-dimensional biplot representation shows only 44% of the cumulative variation, but patterns explaining the additional variation are available from individual axis-loading factors (Table 3). The number of fish in the school was one of the most important driving variables of the first eigenaxis: cumulative fit (cfit) as a fraction of variance = 0.59; the biggest schools of fish in these samples were composed of small Atlantic Menhaden alone or mixed with their predator, Bluefish. The biggest fish, however (typically Bluefish), tended to be alone; thus, another important factor of the first eigenaxis was mean (or Table 2. Definition of categories of fish taxa/ecological groups identified from DIDSON files during preliminary groundtruthing efforts in the Rutgers University Marine Field Station boat-basin and the Hudson River. These categories do not require identification of individual species, but do provide a standardized characterization of fishes for detailed analysis. Category Definition Benthic fish Benthic, often finning but not swimming, no size limit, not oriented to each other Small aggregation of small pelagic fish Swimming in water column, <50 individuals, average length <250 mm Large aggregation of small pelagic fish Swimming in water column, >50 individuals, average length <250 mm Small school of small pelagic fish Swimming together in water column, <50 individuals, average length <250 mm Large school of small pelagic fish Swimming together in water column, >50 individuals, average length <250 mm Single small pelagic fish Swimming in water column, <250 mm Single large pelagic fish Swimming in water column, >250 mm School of large pelagic fish Swimming together in water column, average length >250 mm Unidentified pelagic fish Single fish in water column, measurement not possible Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 201 individual) fish length (cfit = 0.58), but this factor was highly inversely correlated with school size (Table 3, Fig. 3).Variables relating to school organization and the ratio of the imaged body depth to imaged length (aspect ratio), explained variation on the second axis (Fig.3). The mean distance of a fish to its nearest neighbor (cfit to axis 2 = 0.69), was highly correlated to the distance to the next nearest neighbor (cfit = 0.66, Table 3), and reflected the fact that deep-bodied Atlantic Menhaden (which present a relatively high aspect ratio) had a regular spacing among individuals, yet were not always oriented in the same direction. In contrast, Atlantic Silversides, also with uniform (but less) space between individuals, tended to swim as a group with high polarity in organization (cfit = 0.48). Given the sonar tilt, we did not attempt to match the measured aspect ratio of the echogram image with how it translated for the actual fish. The spacing among individuals of an event had a strong effect on both eigenaxes, but it was highly conditional on the presence Figure 3. DIDSON images of various fish schools from the Hudson River study area including: a) large (>50) organized school with 200+ individual fish (4–9 cm in length) near pilings at Pier 40, b) large (>50) unorganized aggregation of fish (6–10 cm in length) near Pier 40, c) small (less than 50) organized school of 40+ fish (7–9 cm in length) near pilings at Pier 57, and d) large aggregation of 200+ fish (8–10 cm in length) mixed with bubbles at the edge of Pier 40. Northeastern Naturalist 202 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 of predators, i.e., schools of Atlantic Silverside were scattered by Bluefish when an aggregation of mixed species was viewed as an event, whereas they could be slightly organized to highly organized when they could be viewed continuously. By themselves, Bluefish were highly unorganized, with individuals appearing to move Table 3. Cumulative fit per variable as fraction of variance in PCA of groundtruthing data.Values indicate the importance of specific measures in resolving the position of different fish classes along a particular axis, and the frequency fitted indicates the relative importance of each axis. Variable Axis 1 Axis 2 Axis 3 Axis 4 Frequency fitted 0.2932 0.1502 0.1387 0.1221 Number of individuals 0.5861 0.6009 0.6408 0.6460 Orientation 0 (two or fewer) 0.1468 0.3101 0.3775 0.5100 Highly ordered 0.0723 0.4819 0.5144 0.7242 Moderately ordered 0.0008 0.0165 0.5203 0.6182 Slightly ordered 0.0880 0.2636 0.2844 0.9110 Random 0.2042 0.2744 0.4707 0.6093 Mean fish length 0.5775 0.5894 0.7778 0.8299 Mean fish height 0.4770 0.7162 0.7168 0.8232 Mean fish aspect ratio 0.0005 0.2746 0.7072 0.7238 Distance to nearest neighbor 0.5076 0.6916 0.7581 0.7654 Distance to next nearest neighbor 0.4472 0.6632 0.7088 0.7366 Depth of fish or school 0.4101 0.4377 0.5082 0.5526 Figure 4. Principle components analysis (PCA) of fish classification from DIDSON images of schooling fish as groundtruthed by net sampling and visual observations at the Rutgers University Marine Field Station boatbasin in Great Bay near Tuckerton, NJ. Menidia menidia (Atlantic Silverside) were common in groundtruth samples and exhibited a wide size range and suite of organization characteristics, but could be readily discerned from deeper bodied fish such as Brevoortia tyrannus (Atlantic Menhaden) and Pomatomus saltatrix (Bluefish). Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 203 independently of other visible fishes; this behavior, also expressed as a low number of individuals per event (making them occur as singletons), served to differentiate them in the PCA. Atlantic Silverside schools took on many different school forms based on the time of day and disturbance from predatory Bluefish or groundtruthing activity as well as from unknown factors. These conditional factors altered relationships seen in the first and second axes so that the major drivers could be important but differently correlated on the third and fourth eigenaxes, a reason for the poor 2-dimensional resolution of the PCA biplot (Table 3). In the Hudson River, other approaches by image reviewers helped differentiate some species. For our identification of large fish, both as schools and individuals (Fig. 5a, b), we relied on characteristics such as caudal fin shape, number of dorsal fins, and shape of pectoral fins, as well as swimming behavior. In many cases, large Figure 5. DIDSON images in the vicinity of piers in the Hudson River. The numbers along the side of each image indicate the distance (m) from the sensor. The images include: a) a school of unidentified large fish (25–30 cm in length) between 4–5 m from the sensor, and acoustic shadows of the same fish at 5–6 m from the sensor; b) a small school of large (30–40 cm in length) fish at approximately 2–2.5 m from the sensor, with a blurred acoustic shadow of the same fish at 5–6 m; c) a school of large (30–40 cm in length) Striped Bass (Morone saxatilis). Note that the dorsal view of the individual (65–70 cm in length; indicated by white oval) at 4 m from the sensor provides a view of the pectoral fins; and d) a single large (70–80 cm in length) Striped Bass (indicated by white oval) at approximately 5 m from the sensor with both dorsal fins evident. Northeastern Naturalist 204 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 Morone saxatilis (Walbaum) (Striped Bass) were readily identified in DIDSON sonograms by the presence of two separate dorsal fins, shallow forked tails, and thick bodies when seen in dorsal view (Fig. 5c, d). In some cases, large (≈1 m TL) Bluefish were distinguishable from Striped Bass based on profile. Identification was aided, in some instances, by the acoustic shadows cast by structural components of the piers (e.g., pilings, Fig. 1) and of the fishes (Fig. 5). Nonetheless, DIDSON file reviewers were often unable to confidently identify other individuals that may have been Striped Bass or other long perciform fishes, such as Bluefish or Cynoscion regalis (Bloch and Schneider) (Weakfish), largely because they were actively swimming and quickly left the field of view. Additionally, while actively swimming, all 3 of these perciform fishes lower their first (spiny) dorsal fin and fold their forked caudal fin to a narrower profile, making species-specific identification difficult. Further, all of these species commonly overlap in size range and timing of occurrence in estuaries. As a group, these 3 species would have been distinguishable from some other perciformes that may occur in the estuaries (e.g., Pogonias cromis L. [Black Drum]), based on shape or maximum size, or both, but we did not observe fishes of such distinguishable forms. Reviewers identified several individuals to a type that suggested either Black Sea Bass or Tautog on the basis of stiff-bodied (pectoral sculling) swimming movements and contact with the bottom as seen during visual groundtruthing. These species and small Cunner could also be difficult to detect by DIDSON if they did not move. In another instance, an Anguilla rostrata (LeSueur) (American Eel) was clearly visible to reviewers, but this species was not detected by any of the groundtruth sampling, although it has been captured previously at the Hudson River study site (Able and Duf fy-Anderson 2005). Discussion Prior studies with DIDSON techniques have demonstrated both advantages and disadvantages of the technology, which vary with type of deployment (static vs. mobile), orientation of the sensor, and range settings at deployment. Most previous attempts to assess fish abundance, biomass, and avoidance (Becker et al. 2011a, Boswell et al. 2008, Crossman et al. 2011, Han and Uye 2009, Holmes et al. 2006, Kimball et al. 2010, Xie et al. 2008) were made from a static location (but see Able et al. 2013, Handegard and Williams 2008) with few attempts to identify faunal components. Visual surveys, as approximated by DIDSON’s sonar-to-image processing, may also be biased against some types of fishes by behavioral traits that limit their detection, such as swimming immediately under the surface, associating closely with the bottom, or by patterns of movement or burial. These same issues have also been identified as problems for visual or video census (Bortone et al. 2000; Sullivan et al. 2000, 2003). One major concern when using DIDSON is the potential for counting-bias. Sedentary species (or those active, but with orientation to a small core area) are best counted by transect designs, but very active swimmers are best counted by fixedradius or point-count surveys (as described by Petit et al. [1995] for forest birds). Also, benthic species may be cryptic in complex backgrounds, either because of a Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 205 loss of color and pattern information with DIDSON, or because they actually hide. Therefore, we focused only on pelagic species when surveying fish response to pier shading with DIDSON (Able et al. 2013). Another shortcoming has been DIDSON’s inability to detect small (less than 40 mm) fishes (Boswell et al. 2008) relative to traditional capture techniques, or to effectively separate these from bubbles, acoustic noise, and non-fish particulates. These problems can be resolved with higher-resolution exchangable auxiliary lenses that concentrate or spread the beams at 1, 3, 8, or 28 degrees (J. Dorsey, Ocean Marine Industries, Inc., Chesapeake, VA, pers. comm.). Finally, as with other acoustic techniques, studies employing DIDSON benefit greatly from the sometimes labor-intensive practice of groundtruthing (Brehmer et al. 2006). Despite these shortcomings, the advantages of DIDSON are numerous. This study extends DIDSON’s utility by demonstrating the effectiveness of surveying fauna from a mobile platform, pinpointing characters that can enhance the identification of fishes, and evaluating reviewer effects on image analysis. DIDSON images can be generated with similar clarity during the day and night, and regardless of the degree of turbidity in the water column, at least in the two estuaries studied here. DIDSON mounted on a kayak provides exceptional mobility, even in very shallow water, something that is not possible with most powered-boat platforms. Additionally, DIDSON does not disturb or attract fauna, as the light necessary for typical video imaging often does. We have seldom observed a fright response by larger predators or smaller prey fish while within the range of DIDSON (typically 1.5–7 m), and have witnessed predation and other social interactions that suggest fish were unconcerned with our presence. In the only other study of this type, which utilized a motorboat, fish exhibited a disturbance response (Xie et al. 2008). In addition, DIDSON software provides the possibility of making fish counts and length measurements from collected images (see above Discussion for caveats). For example, in a South African estuary, researchers used standard DIDSON metrics—size, abundance, and schooling behavior—to infer ecological function relative to turbidity (Becker et al. 2011a) and day–night behavior (Becker et al. 2011b). Further, images of the undisturbed fauna taken from a kayak allow continuous observations that provide details regarding fish schools, shoals, and aggregations (Pitcher 1983), their surface area (Brehmer et al. 2006), and the morphology of individual fish, including the shape of the body and head, the number of dorsal fins, and shape of caudal fins for larger fish. This information is especially helpful because large fish are the most challenging to sample, due to their relative rarity. In addition, continuous imaging provides aspects of swimming behavior, such as mode of propulsion (e.g., carangiform, anguilliform, fin locomotion), degree of schooling, and predator–prey interactions, which can aid identification and give insights into ecological interactions. Thus, DIDSON use is similar to camera use with image–recognition software (Lee et al. 2008, Rova and Dill 2007, Strachan 1993), without the benefit of color. Like video, the data produced can either be enhanced by movement, in the case of static backgrounds, or distorted by the movement of artifacts in moving backgrounds. Northeastern Naturalist 206 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 The dynamics of motion are difficult to express; thus, experience is important in reducing error associated with any DIDSON fish-census from a moving platform. A number of potential measurements taken from static DIDSON freezeframes relate to movements as derivatives (such as measuring distance among nearest neigbors in a school) and can therefore be codified into multivariate ordinations such as PCA. However, even a PCA on our visually groundtruthed data set described considerable overlap among fish, and much of this related to motion and changes of school structure based on the presence of predators. Directly quantifying movement is possible and even routine, as evidenced by the tools provided within the DIDSON software package (DIDSON 2006), but it becomes much more difficult with the currently available software when the background moves relative to the viewer. Thus, the PCA is very useful as a communication tool for quantifying what a reviewer sees in a static DIDSON image, but is not yet a tool for use in automation in a mobile application. Although we have discussed the disadvantages of a heterogeneous and moving background, we note that this belies an important advantage that DIDSON brings to fish surveys, i.e., the ability to simultaneously sample physical habitat structure (Fig. 1). There are few other sampling methods that can do this. Movement also helped in classification, because, unlike a static deployment, it offered different perspectives. One example is in viewing bubbles, which can be confused with small fish. Bubbles, like fish, move in both the horizontal and vertical dimensions because of movement of the kayak and because even very weak currents displace them, and so are difficult to discern, even when using the classic static echogram (time on the x-axis) DIDSON software feature. Bubbles, however, do not have tail beats, which are evident during playback, with even fairly small fish. Future improvements and applications The DIDSON approach, used here to focus on piers and other shallow water habitats, can be applied to other complex urban and natural shoreline habitats, and classification schemes can be constructed and their confidence quantified by extension of the methods presented herein. For example, canonical variates analysis (CVA) uses the same multiple regression algorithm, but combines a known data set with a previously identified data set so that the percent correct identifications of a test set can be quantified. These approaches, including the groundtruthing techniques we have applied in this study and those of others using static DIDSON deployments (e.g., Becker et al. 2011a, b), offer the possibility of determining total juvenile and adult fish-abundance (to some lower size limit), biomass, species or categories composition (with further experience), and selected behavior (schooling, predator–prey, and other social interactions). Together, these approaches can provide parameter estimates during either day or night, which has seldom been accomplished in previous aquatic studies in natural or urbanized habitats (Able and Fahay 2010). Future improvements will further expand DIDSON capabilities when coupled with enhanced groundtruthing (which may need to be location- and habitatspecific for benthic and pelagic fish). Continued development of a static-image Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 207 and movie-clip reference library will expedite the applicability of DIDSON techniques. Enhanced processing power will expand the size range and other capabilities of this system, e.g., higher processor speed will allow tail-beat counts in naturally turbid or dark environments (Mueller et al. 2010). DIDSON recordings from transects can now be merged into mosaics using image landmarks as common reference points without the use of external spatial referencing. Finally, emerging competitive products offering different resolution and pricing will encourage further development and niche-application diversification, including large and small robotic vehicle mounts, diver units with mask displays, and integration with acoustic telemetry technology that identifies acoustically tagged fish with a screen icon (Larry Egan, Lotek Wireless Inc., St. Johns, NL, Canada, pers. comm.). All of these improvements could further expand our understanding of fishes and turbid underwater habitats. Acknowledgments Numerous individuals assisted in these efforts, including summer interns Merrill Rudd and Iris Kemp and volunteers Pat Filardi and Steve Zeck. Technicians from the Rutgers University Marine Field Station—Tom Malatesta, Jamie Caridad, Jen Smith, and Matt Taynor—participated in the field sampling in sometimes trying conditions during the day and night. Franny Buderman conducted preliminary analysis of Pier 57 data as part of her NSF-REU at Rutgers University Research Internships in Ocean Science (RIOS) internship. Hudson River Park Trust personnel assisted with the permitting process. Four anonymous reviewers provided helpful comments on this manuscript, as did the manuscript editor, Thomas J. Maier. Funding for this project was provided by the Hudson River Foundation. This paper is Rutgers University Institute of Marine and Coastal Science Contribution No. 2013-6. Literature Cited Able, K.W., and J.T. Duffy-Anderson. 2005. A synthesis of impacts of piers on juvenile fishes and selected invertebrates in the Lower Hudson River. Rutgers University, Institute of Marine and Coastal Sciences Technical Report #2005-13, New Brunswick, NJ. 38 pp. Able, K.W, and M.P. Fahay. 2010. Ecology of Estuarine Fishes. Johns Hopkins Press, Bal - timore, MD. 566 pp. Able, K.W., J.P. Manderson, and A.L. Studholme. 1998. The distribution of shallow-water juvenile fishes in an urban estuary: The effects of man-made structures in the lower Hudson River. Estuaries 21(4b):731–744. Able, K.W., K.M.M. Jones, and D.A. Fox. 2009. Large nektonic fishes in marsh-creek habitats in the Delaware Bay estuary. Northeastern Naturalist 16(1):27–44. Able, K.W., J. Dobarro, and A.M. Muzeni-Corino. 2010. An evaluation of boat-basin dredging effects: Response of fishes and crabs in a New Jersey estuary. North American Journal of Fisheries Management 30:1001–1015. Able, K. W., T.M. Grothues, and I.M. Kemp. 2013. Fine-scale distribution of pelagic fishes relative to a large urban pier. Marine Ecology Progress Series 476:185–198. Arnold, S.J. 2003. Too much natural history, or too little? Animal Behavior 65:1065–1068. Northeastern Naturalist 208 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 Vol. 21, No. 2 Becker, A., A.K. Whitfield, P.D. Cowley, J. Jarnegren, and T.F. Naesje. 2011a. An assessment of the size structure, distribution, and behavior of fish populations within a temporarily closed estuary using dual-frequency identification sonar (DIDSON). Journal of Fish Biology 79:761–775. Becker, A., P.D. Cowley, A.K. Whitfield, J. Jarnegren, and T.F. Naesje. 2011b. Diel fish movements in the littoral zone of a temporarily closed South African estuary. Journal of Experimental Marine Biology and Ecology 406:63–70. Bortone, S.A., M.A. Samoilys, and P. Francour. 2000. Fish and macroinvertebrate assessment methods. Pp. 127–164, In W. Seaman, Jr. (Ed.). Artificial Reef Evaluation. CRC Press, Boca Raton, FL. 264 pp. Boswell, K.M, M.P. Wilson, and J.H. Cowan. 2008. A semi-automated approach to estimating fish size, abundance, and behavior from dual-frequency identification sonar (DIDSON) data. Journal of Fisheries Management 28:799–807. Brehmer, P., F. Gerlotto, J. Buillard, F. Sanguinéde, Y. Guénnegan, and D. Buestel. 2003. New applications of hydroacoustic methods for monitoring shallow-water aquatic ecosystems: The case of mussel-culture grounds. Aquatic Living Resources 16:333–338. Brehmer, P., T.D. Chi, and D. Mouillot. 2006. Amphidromous-fish-school migration revealed by combining fixed sonar monitoring (horizontal beaming) with fishing data. Journal of Experimental Marine Biology and Ecology 334:139–150. Brown, L.M., K.J. Magowan, D.A. Fox, and J.E. Hightower. 2007. Comparison of splitbeam and hydroacoustic gears for conducting sturgeon surveys. Presented at Mid- Atlantic Chapter of the American Fisheries Society Meeting, Lewes, DE, 1–3 February. Cotterill, F.P.D., and W. Foissner. 2010. A pervasive denigration of natural history misconstrues how biodiversity inventories and taxonomy underpin scientific knowledge. Biodiversity and Conservation 19:291–303. Crossman, J.A., G. Martel, P.N. Johnson, and K. Bray. 2011. The use of dual-frequency identification sonar (DIDSON) to document White Sturgeon activity in the Columbia River, Canada. Journal of Applied Ichthyology 27 (Suppl.2):53–57. Dayton, P., and E. Sala. 2001. Natural history: The sense of wonder, creativity, and progress in ecology. Scientia Marina 65 (Suppl 2):199–206. DIDSON 2006. Fish-assessment selected advanced tools tutorial version 5.07.01 or later, Sound Metrics Corp, Bellevue, WA. Hagan, S.M., and K.W. Able. 2003. Seasonal changes of the pelagic fish assemblage in a temperate estuary. Estuarine, Coastal, and Shelf Science 56:15-29. Hagan, S.M., and K.W. Able. 2008. Diel variation in the pelagic-fish assemblage in a temperate estuary. Estuaries and Coasts 31:3–42. Han, C-H., and S-I. Uye. 2009. Quantification of the abundance and distribution of the common jellyfish Aurelia aurita s.l. with a dual-frequency identification sonar (DIDSON). Journal of Plantkon Research 31(8):805–814. Handegard, N.O., and K. Williams. 2008. Automated tracking of fish trawls using DIDSON (dual-frequency identification sonar). ICES Journal of Marine Sc ience 65:636–644. Holmes, J.A., G.M.W. Cronkite, H.J. Enzenhofer, and T.J. Mulligan. 2006. Accuracy and precision of fish-count data from a dual-frequency identification sonar (DIDSON) imaging system. ICES Journal of Marine Science 63:543–555. Johnston, R., and M. Sheaves. 2008. Cross-channel distribution of small fish in tropical and subtropical coastal wetlands is trophic-, taxonomic-, and wetland-depth-dependent. Marine Ecology Progress Series 357:255–270. Northeastern Naturalist Vol. 21, No. 2 K.W. Able, T.M. Grothues, J.L. Rackovan, and F.E. Buderman 2014 209 Kimball, M.E., L.P. Rozas, K.M. Boswell, and J.H. Cowan, Jr. 2010. Evaluating the effect of slot size and environmental variables on the passage of estuarine nekton through a watercontrol structure. Journal of Experimental Marine Biology and Ecology 395:181–190. Lee, D-J., J.K. Archibald, R.B. Schoenberger, A.W. Dennis, D.K.Shiozawa. 2008 Contour matching for fish-species recognition and migration monitoring. Applications of Computational Intelligence in Biology. 122:183–207 Mueller, A., D.L. Burwen, K.M. Boswell, and T. Mulligan. 2010. Tail-beat pattern in dualfrequency identification sonar sonograms and their potential use for species identification and bioenergetic studies. Transactions of the American Fisheries Society 139:900–910. Pavlov, D.S., E.S. Borisenko, and V.M. Pashin. 2009. Investigations of spawning migration and assessment of abundance of the Kamchatka Steelhead (Parasalmo mykiss) from the Utkholok River by means of DIDSON dual-frequency identification sonar. Journal of Ichthyology 49(11):1042–1064. Petit, D.R., L.J. Petit, V.A. Saab, and T.E. Martin. 1995. Fixed-radius point-counts in forests: Factors influencing effectiveness and efficiency. Pp. 49–56, In C.J. Ralph, J.R. Sauer, and S. Droege (Eds.). Monitoring Bird Populations by Point Counts. USDA Forest Service General Technical Report PSW-GTR-149. USDA Forest Service, Southwest Research Station, Albany, CA 181 pp. Pitcher, T.J. 1983. Heuristic definitions of shoaling behavior. Animal Behaviour 31:611–613. Rova, G.M., and L.M. Dill. 2007. One fish, two fish, butterfish, trumpeter: Recognizing fish in underwater video. In IAPR Conference on Machine Vision Applications. Pp 404–407. Soares Filho, W., J. Manoel de Seixas, and L.Pereira Caloba. 2001. Principal component analysis for classifying passive sonar signals. International Symposium on Circuits and Systems, Sydney, Australia, 2:592–595. Stoecker, R.R., J. Collura, and P.J. Fallon, Jr. 1992. Aquatic studies at the Hudson River Center Site. Pp. 407–427, In C.L. Smith (Ed.). Estuarine Research in the 1980s. The Hudson River Environmental Society Seventh Symposium on Hudson River Ecology, State University of New York Press, Albany, NY. Strachan, N.J.C. 1993. Recognition of fish species by colour and shape. Image Vision Computing 11(1):2–10 Sullivan, M.C., R.K.Cowen, K.W. Able, and M.P. Fahay. 2000. Spatial scaling of recruitment in four continental-shelf fishes. Marine Ecology Progress S eries 207:141–154. Sullivan, M.C., R.K. Cowen, K.W. Able, and M.P. Fahay. 2003. Effects of anthropogenic and natural disturbance on a recently settled continental-shelf flatfish. Marine Ecology Progress Series 260:237–253. ter Braak, C.J.F., and P. Smilauer. 2002. CANOCO Reference manual and CanoDraw for Windows user guide: Software for Canocoal Community Ordination (version 4.5). Microcomputer power, Ithaca, NY. Tiffan, K.F., D.W. Rondorf, and J.L. Skalicky. 2004. Imaging fall Chinook Salmon redds in the Columbia River with a dual-frequency identification sonar. North American Journal of Fisheries Management 24:1421–1426. Turhan-Sayan, G. 2005. Real time electromagnetic-target classification using a novel feature-extraction technique with PCA-based fusion. IEEE Transactions on Antennas and Propagation 53(2):766–776. Xie, Y., C.G.J. Michielsens, A.P. Gray, F.J. Martens, and J.L. Boffey. 2008. Observations of avoidance reactions of migrating salmon to a mobile survey vessel in a riverine environment. Canadian Journal of Fisheries and Aquatic Sciences 65:2178–2190.