Can a Rapid Underwater Video Approach Enhance the Benthic Assessment Capability of the National Coastal Condition Assessment in the Great Lakes?

The target population was defined by a GIS frame to represent the coastal water resource of the Great Lakes. Briefly, the nearshore zone was defined as a narrow zone from the shoreline (buffered to include river mouths and smaller indentations with waters continuous with the lake) to the 30-m-depth contour, unless it first reached a maximum of 5 km from shore. The entire nearshore survey frame, the GIS-explicit approximation of the target resource, covered 18,881 km² (12 %) of the U.S. Area of the five Lakes. The 2010 NCCA was conducted for the whole U.S. Great Lakes coastal zone, with an approximate shoreline length of 6100 km, measured as a simplified linear waterline. The sample population for the survey was about 45 nearshore sites per lake; with 10 % of the sites revisited within the summer season. Sample collection occurred from May to September 2010. The embayment frame was a subset of the nearshore defined by geometric size and shape criteria to represent a target population of area close to shore and vulnerable to landscape-derived stressors, with more restricted water circulation and, as a population, shallower water than the more physically open nearshore. Embayment polygons were semi-enclosed (geomorphically) areas in small to medium harbors and bays from ~1 to <100 km². Larger bays like Green Bay and Saginaw Bay had sites as part of the nearshore frame. The embayment area was 929 km², or 4.9 % of the nearshore zone, and sampling included 151 sites across the Great Lakes region. NCCA site locations for both nearshore and embayment populations were selected as a probability survey, using generalized random tessellation stratification (Stevens and Olsen 2004). Additionally, as an extension to the NCCA, sites were also video sampled by partners in National Park Service (NPS) waters. In evaluations of video performance, 362 videos were used (nearshore sites, embayment sites, NPS sites, and revisited sites). When comparing video findings directly to estimates based on PONAR data, we only used 2010 NCCA design sites: 309 videos.

A total of 362 videos were collected, representing 78 % of all sites visited in the Great Lakes for the NCCA survey in 2010: 128 in Lake Michigan, 87 in Lake Superior, 48 in Lake Huron, 43 in Lake Erie, and 56 in Lake Ontario. 55 % of all videos had a video quality rating of excellent, 35 % marginal, and 10 % poor.

Not all videos had the quality necessary for confident use. The reason for rating videos marginal or poor was classified as controllable or uncontrollable. 32 % of the marginal and poor videos had controllable reasons for their rating: (1) the view was not close enough to the bottom and/or rocks (86.8 %) (2) the camera light needed adjustment (7.5 %) (3) the camera angle was poor (3.8 %) (4) the camera moved too fast (1.9 %). Adjusting deployment procedures to correct for these controllable reasons could yield more usable images. In general, high current flows were not often encountered or mentioned in field notes. Currents could be additional complicating factors in areas within connecting channels (not sampled in 2010) or at the edges of selected river mouths. Such situations might be considered controllable by technique modifications (drifting with the current, deploying on a weighted bottom sled, etc.), but they were not further considered here.

Conversely, 68 % of the marginal and poor videos were classified as having uncontrollable reasons for their rating: (1) the image was turbid or had a dark green or brown tint (36.9 %) (2) the site had heavy vegetation which blocked the view or increased particulates when the vegetation was agitated by the camera (25.2 %) (3) the Goby-like fish did not display themselves in a way to confidently see a dorsal spot (18.9 %) (4) the video had other infrequent issues such as mechanical problems, waves causing the image to be too bouncy, or sunlight rendering the image too bright (18.9 %). The most frequent problem of turbidity and dark tints is a common difficulty with video sampling (Van Rein et al. 2009). Researchers have found it difficult to detect small organisms like the Round Goby in their video images when water quality parameters were poor (Schaner et al. 2009; Wilson et al. 2006). Advances in technology have been made to improve imaging in turbid conditions (Spink and Read 2011). Whether such technology, or something similar, is feasible and cost effective for the NCCA needs to be determined.

To further examine the relationship between video performance and water clarity, video clarity ratings were compared with K_d values to identify what specific conditions result in a marginal or poor image. The Kruskall–Wallis test showed that mean ranks of K_d values per site were significantly different among excellent, marginal, and poor video images (H = 63.999, 2 df, P < 0.05). K_d values were significantly higher (less light penetration) at sites where marginal (Q = 5.081, 2 df, P < 0.05) and poor (Q = 7.030, 2 df, P < 0.05) images were collected. In addition, K_d values were significantly higher at sites with poor images compared to sites with marginal images (Q = 3.324, 2 df, P < 0.05). Examining the histogram and means (Fig. 2) of K_d values across the three ratings indicated a possible threshold for capturing an excellent video was approximately K_d ≤ 0.5/m. Logistic regression further enforced that video clarity was inversely related to K_d (X² = 59.007, df = 1, P < 0.05). The logistic model predicted that a K_d = 0.5/m would give a 58.9 % chance of capturing an excellent video (Fig. 3).

Chlorophyll a concentration at sites was examined to further explore video image clarity. The Kruskall–Wallis test showed that mean ranks of chlorophyll a concentration per site were significantly different among excellent, marginal, and poor video images (H = 81.830, 2 df, P < 0.05). Chlorophyll a concentrations were significantly higher at sites where marginal (Q = 6.354, 2 df, P < 0.05) and poor (Q = 7.764, 2 df, P < 0.05) images were collected. In addition, chlorophyll a concentrations were significantly higher at sites with poor images compared to sites with marginal images (Q = 2.724, 2 df, P < 0.05). The histogram and means (Fig. 4) of chlorophyll a concentrations across the three ratings indicated a possible threshold for capturing an excellent video was approximately 3.5 µg/L. Results of the logistic regression further supported that video clarity was related to chlorophyll a concentration (X² = 99.231, df = 1, P < 0.05). The model showed that a 3.5 µg/L chlorophyll a concentration would result in a 53.5 % chance of capturing an excellent video (Fig. 5).

Examining how K_d and chlorophyll a concentration jointly influenced video clarity at sites showed how the ability to capture an excellent video image was influenced by water quality parameters related to clarity (Fig. 6a). Most excellent videos are associated with low K_d values and chlorophyll a concentrations. Application of both proposed threshold values encompasses the majority of excellent images as seen by the shaded region (Fig. 6a, b). Four sites with excellent images were not included in the shaded region and were outside the threshold of both K_d and chlorophyll a concentration (Fig. 6a). Sites within the threshold bounds highlight the need to acknowledge the effect of environmental limitations on capturing high quality videos (Fig. 6b). To further illustrate such constraints, we estimated the average values for chlorophyll a and K_d for the whole set of NCCA sites as 3.84 µg/L and 0.33/m, respectively; cumulative distribution functions showed that about 10–20 % of the survey area would exceed either threshold. Evaluation of the embayment sub-resources as distinct from the nearshore yielded average embayment values significantly higher than the nearshore alone. Means were 4.96 µg/L and 0.68/m for chlorophyll a and K_d, respectively, and thus above the identified video thresholds. Cumulative distribution functions showed that about 30–50 % of the embayment area (only <5 % of the nearshore size) exceeded either threshold.

In total, 49 sites that were not sampled by PONAR had an excellent video captured. Approximately 73 % of these sites had visible rock substrate, which likely prevented a successful PONAR grab. The ability to sample hard bottom is an important piece of the coastal condition assessment puzzle, which is often lost and prevents us from seeing the whole picture. Invasive dreissenid mussels have been found in higher densities on substrates such as bedrock, cobble, and wreck debris (Coakley et al. 1997).

Video sampling detected dreissenids at eight sites where PONAR sampling did not, and at 17 sites were PONAR sampling was unsuccessful. PONAR sampling may have missed dreissenid detection at those 8 sites if their distribution was patchy, or if they were found solely on hard substrates (Coakley et al. 1997). At 32 sites video sampling did not detect dreissenids even though PONAR data showed they were present. Half of these 32 sites had less than 5 dreissenids per PONAR grab. Dreissenid abundance does influence video detection ability (X² = 14.576, df = 1, P < 0.05). The logistic regression model predicted a 34 % chance of dreissenid detection from video sampling when abundance in a full size PONAR was 5 (Fig. 7), which is equivalent to ~96 mussels/m². As mussel abundance increases, so does the video’s mussel detection ability (Fig. 7).

Deciding what sampling gear to use in an ecological study is one aspect that can affect the scope of the resulting data and in turn the overall findings of the study. For the NCCA, benthic grab sampling offered data regarding sediment type, color, and odor. In addition, the sample was processed to identify benthic macroinvertebrates that were used as indicators for ecosystem health. While this small-scale sampling does provide good large-scale information, one must recognize the limitations of a grab sample fully representing the entirety of a site. Vegetation, rock, and organisms can be present in patches, and therefore could be easily missed by such a sampling technique. Toward this point, we make a brief assessment using mussels as our example to show how video affected quantitative estimates of percentage area coverage by mussels.

A strength of the NCCA survey, with its explicit spatially balanced statistical design foundation, is that it can be used to estimate the fraction of a total resource in a certain condition. In this case, we estimated the portion of the total area surveyed that had populations of live dreissenid mussels. For the whole U.S. nearshore region (including embayments), PONAR sampling estimated that 38 % of the entire zone had dreissenids. When all successful video samples were used to supplement the PONAR sampling, the estimate rose to 44 %. The estimate represented an area with mussels present as 7570 km², with a 95 % confidence interval of ±7 %. The increase with supplemental video was significant at p = 0.057 (one-sided Z test). This difference is, quite likely, conservative. As we noted, there was a detection limit for video when dreissenids are present in low concentrations. Moreover, we need to strive to improve controllable video performance, as more quality results might also slightly elevate estimates. Related to this, there were some sites where a few mussel shells were noted but we did not have enough confidence in the images to categorize as having presence of live mussels. Overall, video results raised regional nearshore average estimates by 15 % and significantly more than that in a couple lakes (Huron and Ontario).

Figure 8 graphically compares results, with and without supplementary video results added to PONAR-based estimates, for the whole region’s nearshore and embayments, as well as for the nearshore of each individual lake. This illustration is not to show a correlation or regression between the two sets—they are not independent. Instead it simply highlights that in most cases video inclusion into the overall assessment raised the area estimate, i.e., points mostly fall well above a 1:1 line. Thus, video sampling tended to increase the estimated area occupied by dreissenids compared to PONAR sampling alone. Lakes Erie, Ontario, and Michigan showed the highest occupancy (54–59 %, with video included). Lake Superior showed the lowest percent occupancy (by PONAR and video). Even though we supplemented many clear-water Lake Superior sites that were not PONAR-able with videos at a number of rocky-bottom sites, the area with dreissenid mussels present did not increase because dreissenids were not present on hard substrate. Three of the four Lake Superior sites where dreissenids were detected were soft-bottom sites in the Duluth-Superior Harbor area (a known zebra/quagga area, Grigorovich et al. 2008). The only other site with a dreissenid mussel (a sandy nearshore environment eastward of Chequamegon Bay WI) had only one specimen identified in a PONAR sample.

Therefore, PONAR-based surveys will slightly underestimate the actual extent of dreissenid mussels in many, but not all, nearshore areas. The embayments as a whole (as a sub-population within the nearshore) had a slightly higher presence of mussels. Video supplement was less effective for embayments in the sense that it did not significantly increase the area-occupied estimate (48 vs. 50 % of the embayment sub-area, Fig. 8). As previously noted, embayments as a group had higher K_d (lower water clarity) and thus the success of video was more limited (as per Fig. 3) than in the clearer open nearshore. In contrast, hard substrate was more prevalent in the open, more exposed areas of nearshore, which cannot be sampled by PONAR. We noted above that where PONARs were not taken and video was collected, the large majority of sites showed rocky bottom. Enhancement to benthic condition assessment in the Great Lakes with our video protocol was in this case, and generally will be, more pronounced for the open nearshore than in the backwater, depositional softer bottom sediments more characteristic of embayments.

PONARS and video samples were obtained at a similar proportion of the total sites (76–78 %), but the two sets of successful sites were only partially intersecting. Video was successful at a high proportion of sites without PONAR; therefore the union of their sampling sites was a larger set than that of PONAR alone. The un-surveyed area (by PONAR), 22 % of the total area, had a significant presence of mussels (17 of 30 excellent videos), which otherwise would have gone undetected. On the other hand, mussel detection rate by video was lower where PONARs were collected, because there were more marginal and poor videos due principally to water clarity. Thus, to optimize mussel detection for future NCCA surveys: (1) sampling by video should be a high priority where PONARS are not successful and (2) video methods should consider options available to overcome poor water clarity situations.