Evaluation of Recreational Impacts on Eelgrass Using Unoccupied Aerial Systems and Virtual Ground Truth Data

Eelgrass (Zostera sp.) provides a variety of important ecosystem services including stabilizing sediments, buffering storm surge, and nutrient filtration (McGlathery, Sundback, and Anderson, 2007; Nahirnick et al., 2019; Nordlund et al., 2016). It also provides critical habitat for salmon, crab, shellfish, and seabirds (Bulthuis, 1995; Thom et al., 2003). Recreational impacts from boat traffic (Sagerman, Hansen, and Wikstrom, 2020), docks (Logan et al., 2022), and mooring (Parry-Wilson et al., 2019) have been shown to have adverse impacts on eelgrass. Boat propellers can cut or uproot the vegetation and turbulence from propellers can resuspend sediments, generating turbidity, shading of the vegetation, and smothering as the sediment settles (Burgin and Hardiman, 2011). Chemical pollution from fuel, lubricants, and antifouling paints can also have impacts (Egardt et al., 2018). Docks can affect eelgrass because of shading (Eriander et al., 2017) and altered hydrodynamics, resulting in erosion and sediment translocation (Dugan et al., 2011). Anchoring and mooring create mechanical damage to vegetation and can also stir up sediments (Ostendorp et al., 2009). Two species of eelgrass occupy the inland waters of the Pacific Northwest: Zostera marina and Zostera japonica. Zostera marina, the native and most abundant species, is found in mid- to subtidal regions and Z. japonica, a nonnative species, is less abundant and is found in upper to mid-tidal ranges (Bulthuis, 1995; Kaldy, 2006; Kaldy, Shafer, and Magoun, 2015; Kim et al., 2016). Zostera japonica is believed to have reached North America from Japan in the packing materials of clam exports (Bulthuis, 1995; Kaldy, 2006; Shafer, Kaldy, and Gaeckle, 2013).

Traditional methods for monitoring eelgrass have involved the use of ground-based surveys (Neckles et al., 2012) or towed submerged video (Nahirnick et al., 2019), which are quite time consuming and limited in spatial extent. Imagery from satellites (Forsey et al., 2020; Hogrefe et al., 2014; O’Neill and Costa, 2013) and traditional aircraft has also been used (Clinton et al., 2007; Fletcher, Pulich, and Hardegree, 2009; Neckles et al. 2012; Young et al., 2008). Although ideal for regional-scale studies, this type of imagery can be expensive, has rather low spatial resolution (∼1–30 m), and timing image acquisition to tidal cycles can be problematic. Coordination of image acquisition and collection of ground truth data can also be challenging.

More recently, the use of imagery acquired using low-cost, unoccupied aerial systems (UAS) has been used (Bergner et al., 2025; Duffy et al., 2018; Leblon et al., 2022, Nahirnick et al., 2019). The advantage of UAS imagery includes the low cost, very high spatial resolution (a few centimeters), and ease of timing image acquisition to tidal cycles. The disadvantage of UAS imagery is that the large data volume and computational load associated with high spatial resolution as well as operational constraints (e.g., battery limitations on flight time as well as regulations restricting flights beyond visual range of the operator) currently limit the spatial extent of these studies to a few square kilometers. These previous efforts to monitor eelgrass using imagery from satellites, traditional aircraft, or UAS have not been successful in distinguishing between different species of eelgrass and have primarily been limited to simply mapping the absence or presence of eelgrass. Eelgrass presence is generally defined as >5–10% eelgrass cover. However, one recent UAS base study (Bergner et al., 2025) was successful in distinguishing between Z. marina and Z. japonica and broke out four broad percent cover categories.

The objective of this study was to evaluate the use of simple vegetation indices (VIs), derived from multispectral imagery acquired using UAS, in conjunction with a novel approach for obtaining virtual ground truth data to quantify variation in the percent cover of eelgrass and to evaluate the effect of recreational boat launch activity on eelgrass.

The study site was in Washington State on the Salish Sea, a transnational body of water that includes Puget Sound and the straits of Georgia and Juan de Fuca (Sobocinski, 2022). Wildcat Cove (48°38′59.9994″ N, 122°29′23.9994″ W) is part of Larabee State Park (Figure 1). The cove includes a boat ramp that is heavily used during the summer months, particularly during the recreational crabbing season, which typically opens in mid-July. The park maintains an automated device to record all vehicles entering the parking lot and most, but not all, vehicles that enter the lot are there to launch a boat of some kind (Figure 2). One anomaly with these data is that, during 2023, the traffic counter was incorrectly positioned so that it was triggered by passing trains on a track adjacent to the entry road. The “count” for a passing train represented the time it took for the train to pass. Repositioning the sensor in 2024 eliminated this issue, resulting in more accurate counts. During June and July, the difference between the 2 years averaged 85 cars/d, presumably due to the exclusion of incorrectly recording passing trains during 2023. From 1 June until the opening of recreational crabbing the count averaged 220 and 135 cars/d during 2023 and 2024, respectively. On the first Saturday of recreational crabbing season there were 511 and 415 cars recorded in 2023 and 2024, respectively, although the 2023 record includes the counts of passing trains. After the opening of recreational crabbing, the highest number of visits occurred on the 5 days each week when crabbing is permitted, and the lowest numbers occur on the 2 days each week when state regulation closes the area for crabbing. Despite the anomalous logging of passing trains in 2023, these data document the heavy use of the boat ramp throughout the summer months as well as the large increase that occurs with the opening of the recreational crabbing season.

During low tides, much of the cove is exposed mudflat and vehicles with boat trailers typically back down across the mudflat to the water’s edge. Kayaks and other human-powered boats are typically carried or dragged across the mudflat to the water’s edge, but these users sometimes drive across the mudflat as well. The maximum tidal range for this site during 2023 was 4.16 m (NOAA, 2023) and the maximum water depth at the subtidal eelgrass on the edge of our study area is approximately 5–6 m at the highest tide stage.

Imagery was acquired using a DJI Matrice 210 quadcopter that carried both a DJI X4S RGB camera as well as a Micasense RedEdge MX 10-band multispectral camera (Micasense, 2023; Table 1). Survey flights were conducted using the DJI Pilot app at an altitude of either 40 or 60 m above ground level (AGL) with a sidelap and frontlap of 75%. All flights were conducted in compliance with Federal Aviation Administration Section 107 rules.

Flights were timed to occur at or near low tides and under conditions recommended by Bergner et al. (2025), Hein (2022), and Nahirnik et al. (2019). This included the objective of flying when the sun angle was below ∼52°, wind <5 m/s, and cloud cover <10% or >90%. These previous studies have shown that UAS flights occurring at higher sun angles result in specular reflection off the water surface, resulting in glare that makes the imagery unusable. Bergner et al. (2025) also noted that diffuse illumination that occurs with cloud cover >90% can result in reduced levels of glare that can still be problematic. Higher wind speeds generate a dappled water surface that can also result in scattered glare, even at lower sun angles. Finally, even scattered cloud cover can result in shadowed areas in the imagery. To ensure the best-quality imagery, sun angles for each day of data collection were assessed using NOAA’s (2022) sun angle calculator.

Before each flight, 30 numbered 45 × 45 cm ground control panels were placed along three transects on the exposed mudflat. These transects were spaced ∼15 m apart and, along each transect, the panels were spaced ∼5 m apart. Each transect extended from the upper intertidal to the edge of the subtidal zone. Panels were held in place with small stakes through the center of the panel. Transect placement varied slightly between dates but covered the same general location. Before the survey flight, the UAS was manually flown over each transect at an altitude of 5 m AGL to obtain standard red, green, blue (RGB) images of each panel with the X4S camera. These images provided virtual ground truth data (described below) that was used to model total vegetation percent cover.

After returning from the field, and for each ground control panel image, virtual ground truth data was obtained by superimposing a 4 × 4 grid onto the image next to each panel and resizing the grid to match the size of each panel (Figure 3). Resizing the grid to match the size of the ground control panel ensured that the grid corresponds to a consistent area on the ground. The cover type was visually assessed at the corner of each grid cell, resulting in data for 25 points within this 45 × 45 cm sample grid. Percent cover for each cover type was then calculated for each sample grid on the basis of these sample points. The four cover types included eelgrass (Zostera sp.), algae (Ulva sp., mostly Ulva intestinalis), bare, and detritus. No distinction was made between the two species of eelgrass. Detritus was mostly composed of dead eelgrass. All visual assessments of cover were conducted by the same individual for all four dates. This approach is equivalent to the ground-based point intercept method for the assessment of eelgrass and other cover types described by Clinton et al. (2007). Their approach involves the use of a sample frame that includes a 4 × 4 grid of string to define 25 sample points with the visual assessment done in the field.

Each survey flight resulted in several thousand images that were processed using the Agisoft Metashape Professional structure-from-motion software (v 1.8.4; 64 bit). The coordinates of seven ground control points (GCPs) were located in Google Earth imagery. These GCPs consisted of the edges of large rocks that could be easily located in the imagery and were used to coregister imagery from all four dates to facilitate change analysis. Image processing in Agisoft Metashape resulted in a 10-band orthomosaic image for each flight and these images were exported from Agisoft Metashape as tag image file format (TIFF) files and opened in the ENvironment for Visual Images (ENVI) software (v 5.7) for further processing. Two versions of these TIFFs were generated for each flight: one resampled to 5-cm grid cells and another resampled to 0.5-m grid cells. This larger grid cell size closely matched the size of the virtual ground control plots.

Variables used in this equation are presented in Table 2. Each of these VIs has been widely used and has been found to be a useful predictor of a variety of parameters for land-based studies, including photosynthetic rate (Sellers, 1985; Tucker and Sellers, 1986), chlorophyll content (Barnes et al., 2000), leaf area index (Delegido et al., 2013), biomass, and crop health (Huang, Tang, and Hupy, 2020).

Although the coordinates for each of the virtual ground control panels were not recorded, each panel was easy to locate in the 5-cm-resolution imagery. With the 5-cm-resolution imagery as well as each of the three 0.5-m-resolution VIs loaded in ENVI, each panel was located in the 5-cm-resolution imagery, and the value for each VI adjacent to the panel was recorded from the 0.5-m-resolution imagery. In addition to data from the virtual ground truth plots, for the first three dates, VI values were also collected for eight arbitrarily selected points in the subtidal eelgrass. Inspection of the images for these dates made it clear that these samples represented 100% eelgrass cover that was floating on the water surface in July, but by late August, subtidal eelgrass cover was declining somewhat, so no VI values from subtidal eelgrass for August were recorded. The data for each index from all flights were then compiled and merged with the percent cover data obtained from the virtual ground control plots.

Data for all four dates were combined and each VI was used to develop a regression model of percent coverage of eelgrass. Each VI was also used to develop a regression model for total vegetation cover, consisting of the sum of percent coverage of eelgrass plus percent coverage of algae. This resulted in six separate regression models. Regression analysis was conducted in R (version 4.3.2) using K-fold cross-validation (k = 10) with ordinary least-squares regression analysis (R Core Team, 2022).

Four survey flights were conducted between 14 July and 28 August 2023. Weather conditions and tide stage during each flight are presented in Table 3. The 14 July flight occurred the day before the opening of the recreational crabbing season. The 14 July flight was conducted at an elevation of 60 m AGL and resulted in a ground resolution of 4 cm, and this flight was completed in 15 minutes. All subsequent flights were conducted at an elevation of 40 m AGL, resulting in a ground resolution of 2.5 cm, and these flights were completed in 20 minutes. The manual flight to obtain the virtual ground truth images was conducted before each survey flight and was completed in 10–12 minutes. On each day, the total time required for deployment and retrieval of GCP panels, completion of both the manual and survey flight, as well a battery change between flights was <1 hour. GCP panels were deployed ∼35 minutes before low tide and the manual flight was concluded ∼10–15 minutes before low tide. This made it possible for the survey flight to start just a few minutes before and conclude just a few minutes after the lowest tide on each day. On each day, tide stage was observed to remain within ∼0.1 m of the lowest tide stage for ∼90 minutes.

RGB images generated from the 5-cm resampled versions of the 10-band orthomosaics for each date are presented in Figure 4 and enlarged subsets from each date are presented in Figure 5. Note that a denuded track through the eelgrass on the exposed mudflat was visible on 14 July but that the width of this track had increased and was more deeply rutted on 17 July after two busy days of boat launch activity on the first weekend of crabbing season. A track of reduced eelgrass cover was also visible in the subtidal eelgrass on 14 and 17 July (Figure 4a, b). Eelgrass in the region typically achieves peak biomass in late July (Bulthuis, 1996), and regrowth of the track through the subtidal eelgrass was apparent in the 31 July image (Figure 4c). Much of the decline in the subtidal eelgrass and vegetation cover elsewhere in the cove on 28 August may have resulted from typical seasonal senescence. Data from the virtual ground truth plots were consistent with this seasonal dynamic (Table 4). Eelgrass was the dominant cover type in these plots during July, but by late August, senescence of eelgrass was occurring and detritus, mostly consisting of dead eelgrass, was the dominant cover type in the plots (Table 4). A few of the panels were displaced by wind before the survey flight and could not be used, and this resulted in some variation in sample size among flights.

Selection of days and times for flights requires consideration of both tide stage and weather conditions. The objective was to fly at the lowest possible tide and the lowest possible sun angles. Ideally, flights would occur at identical tide stages, but this is not practical, and some variability is inevitable (Table 3). Sun angle also varied considerably, and this resulted in shadow on the south side of the cove during two flights (Figure 4a,c). Higher wind speeds on 17 July resulted in scattered glare in the open water outside the cove (Figure 4b), but this did not influence the imagery acquired on the primary study area inside the cove. All July imagery was acquired with little or no cloud cover. Although there were some clouds during the 17 July flight, no cloud shadows were present in the study area during the flight. 28 August was 100% overcast but despite the very low sun angle (Table 3) the diffuse illumination resulting from the overcast resulted in no shadows (Figure 4d). Scattering caused by this overcast meant that illumination was coming from all directions, including a portion that was coming from directly overhead. This results in specular reflection and glare, but glare that was not as extreme as would occur with clear skies and the sun directly overhead. The effect of this in portions of the mudflat that retained a few centimeters of water is illustrated in Figure 6. Virtual ground truth plots that exhibited this type of glare on 28 August were not used to model vegetation cover.

The best model was based on the normalized difference index (NDI, Table 2) and explained 82.6% of the variation in percent total vegetation cover (Table 5 and Figure 7). A model based on the use of normalized difference vegetation index (NDVI) explained 81% of the variation in percent total vegetation cover. Separate analyses using just percent cover of eelgrass alone were also significant, but the percent variance explained was only 65.2% or 65.6% when using NDI or NDVI, respectively (Table 5). This reduction in explained variance was due to the noise generated by the presence of algae. The normalized difference red edge performed poorly for predicting either percent total vegetation cover or eelgrass alone.

The relationship from Figure 7 was used to model total vegetation cover in the study area. The area that included the intertidal mudflat and the subtidal eelgrass was manually delineated, and the model was only applied to this area. The results of doing so are presented in Figure 8. Change in total vegetation cover between dates was then quantified by differencing these vegetation cover layers (Figure 9). Between 14 and 17 July, vegetation cover declined by >10% on 44.3% of the study area (Table 6). Vegetation cover increased by >10% on 50.6% of the study area between 17 and 31 July. Vegetation cover again declined by >10% on 70.9% of the study area between 31 July and 28 August.

A clear track with low vegetation cover from the paved boat ramp to the water’s edge and into the subtidal eelgrass was already apparent on 14 July (Figure 8a). Extensive loss of vegetation cover occurred between 14 and 17 July on either side of this main path from the paved boat ramp to the water (Figure 9a). Presumably, this resulted from a busy weekend with multiple vehicles driven across the mudflat to launch boats simultaneously and fanning out across the mudflat. Heavy foot traffic may also have contributed to the observed vegetation loss. Loss of subtidal eelgrass presumably results from damage caused by outboard motor propellers and at higher tide stages, when the mudflat is inundated, this also probably contributed to the loss of eelgrass in the intertidal zone. More limited loss of vegetation cover on the mudflat occurred between 17 and 31 July (Figure 9b). The increases in vegetation cover away from this path during this period may have resulted from the typical seasonal growth of both eelgrass and algae as they reach the peak of their biomass at this time of year (Bulthuis, 1996). There also appears to have been a modest reduction in recreational activity after the first week of crabbing season (Figure 2). Also note the regrowth of subtidal eelgrass in the area that showed loss during the previous period. There were additional reductions in vegetation cover along the boat launch track between 31 July and 28 August (Figure 9c), but the extensive declines in vegetation cover throughout the cove may result from seasonal declines in both eelgrass and algae. Note the large mat of dead eelgrass rafted up on the upper intertidal zone in Figure 4d. However, the 28 August results should be interpreted cautiously since the 100% overcast conditions and resulting glare in areas of shallow water (Figure 6) may have resulted in spurious predictions of vegetation cover loss.

These results demonstrate that a simple VI derived from UAS imagery, combined with the use of virtual ground truth data, can be used to quantify the percent coverage of eelgrass and algae in the intertidal and upper subtidal zone. This approach was successful in documenting both recreational impacts on vegetation cover as well as seasonal growth and senescence. The novel approach, described here, involving the use of virtual ground truth data, minimizes the amount of time required in the field compared with the use of traditional ground-based sampling. As described above, it took less than an hour in the field each day to complete all data collection tasks. This included deployment and retrieval of the GCP panels, completion of the manual flight to obtain the virtual ground truth imagery, the survey flight, and a battery change between flights. This made it possible to complete all fieldwork within a very short window of time centered on the lowest tide stage. The process of visually assessing the ground truth imagery to quantify the occurrence of each cover type adjacent to each ground control panel was then completed later in the office. Doing this assessment in the field using a sampling frame containing a grid of string, as described by Clinton et al. (2007), would add several minutes for each sample location. For the 30 samples used on each day in the current study, this would add an hour or more to the time needed in the field. Additional time savings could be achieved by using a UAS equipped with real-time kinematic (RTK) GPS to obtain the ground truth imagery. RTK-GPS makes it possible to precisely geotag imagery and therefore confidently locate the precise position of the ground truth locations in the lower-resolution ortho image of the entire study area. This would make it possible to eliminate the need to traverse the study area to deploy and retrieve GCP panels. This could result in considerable time savings, particularly for larger study areas. Furthermore, with one UAS programmed to complete the survey flight at a higher altitude (∼40–60 m AGL), a second RTK-GPS-equipped UAS could be programmed to simultaneously fly at a very low altitude (∼5 m AGL) to a set of randomly selected points to acquire ground truth images.

The VIs that proved most useful in the current study (NDI and NDVI) relied on the measurement of reflectance in the near-infrared (IR) and red edge portions of the spectrum (Table 2); however, water absorbs very strongly at these longer wavelengths, especially the near-IR (O’Neill and Costa, 2013). Land-based studies have long established that these longer wavelengths are critically important for remote sensing of vegetation (Delegido et al., 2013; Tucker, 1978). As noted above, a key challenge for eelgrass mapping efforts that are based on the use of data from either satellites or traditional aircraft is timing image acquisition to tide stage. This generally means that imagery is not obtained at the lowest tide stage and much, if not all, of the eelgrass is submerged. For this reason, these studies are limited to utilizing data from the visible portion of the spectrum with this limited information content. It seems likely that this is what has limited these previous studies to simply mapping the presence or absence of eelgrass, albeit with rather high classification accuracy (reviewed by Forsey et al., 2020). It seems likely that the success of the current study in quantifying percent cover of eelgrass and algae was made possible by precisely timing image acquisition to the lowest tide stage when the intertidal eelgrass is exposed and eelgrass in the upper subtidal is floating on the water surface. This facilitated the use of the information-rich near-IR and red edge portions of the spectrum.

The recent study by Bergner et al. (2025) also timed UAS image acquisition flights to the lowest tide stage and focused on eelgrass that was fully exposed in the intertidal zone. They also used the same 10-band camera that was used in the current study. Each of the VIs used in the current study relies on only two of these 10 bands. Rather than using these simple VIs, Bergner et al. (2025) conducted a principal components analysis (PCA) using all 10 bands and then used the first three of the resulting PC axes as predictor variables. The combination of obtaining imagery of exposed eelgrass as well as the use of PC axes that included the near-IR and red edge bands contributed to their ability to distinguish between areas dominated by the native Z. marina vs. the nonnative Z. japonica. They were also able to break out areas containing a mixture of the two species as well as bare areas with sparse eelgrass cover. The use of PCA in the current study may have made it possible to distinguish between eelgrass and algae. However, UAS cameras with ∼10 bands are not widely available. Cameras that include four or five bands, including a red edge and near-IR band, as well as two or three bands in the visible portion of the spectrum, are much more widely available. As such, the current study focused on evaluating the use of simple VIs that can be more easily used by a wide variety of investigators.

This work was motivated by concerns over recreational impacts on eelgrass. However, recreational activity may have broader impacts on a variety of ecosystem services and on other organisms that occupy this habitat, including crabs, clams, and salmon. By focusing on variation in the coverage of both eelgrass and algae, a model with more predictive power was developed (Table 5) and one that may be doing a better job of capturing these broader environmental impacts rather than impacts on eelgrass alone. With additional work, it would be possible to focus more narrowly on eelgrass, as well as distinguish between Z. marina and Z. japonica (Bergner et al., 2025). However, documenting the impacts of boat launch activity on both eelgrass and algae may provide a broader assessment of environmental impacts.

These results clearly indicate that boat launch activities in Wildcat Cove have an impact on the coverage of eelgrass and algae. The spatial extent of the impacts was primarily limited to an area that was roughly 20 × 90 m. On busy weekends, as multiple groups were launching boats, vehicles fan out across the mudflat and the width of the affected area expands. On 17 July, after the busy opening weekend of crabbing season (Figure 2), the main path through the mudflat was deeply rutted and a bit wider (Figures 4b, 5b). It is apparent that some vehicles swing out quite far from the main track either to avoid other vehicles or seeking a less rutted track (Figure 5b–d). It is also apparent that the passage of boats through the subtidal eelgrass results in loss of eelgrass, presumably from boat propellers (Figures 4a,b; 8a,b). By 31 July, the affected area on the mudflat had continued to expand and there were a few stray scars in the eelgrass well off to the side of the main track. Late July is near the peak of the growing season for eelgrass and the track through the subtidal eelgrass appears to have regrown and mostly filled in (Figures 8c, 9b). By 28 August, there appears to be some expansion of impacts along the main launch track across the mudflat, but this is difficult to evaluate since there is an obvious decline in vegetation cover throughout the cove that may simply reflect typical seasonal declines.

These results contribute to an evolving understanding of best practices that are needed to obtain useful imagery with UAS in nearshore marine environments (Bergner et al., 2025; Hein, 2022; Nahirnik et al., 2019). These previous studies have suggested that flights should be conducted at sun angles of <52°, wind <5 m/s, and cloud cover <10% or >90%. These factors combine to reduce the incidence of glare from the water surface, which makes imagery unusable. Bergner et al. (2025) further suggested that flights with >90% cloud cover are to be avoided since diffuse illumination under these conditions results in glare, albeit somewhat reduced glare. Results from the current study confirm this suggestion (Figure 6). Even with a very low sun angle on 28 August (Table 3), the diffuse illumination resulting from the 100% overcast conditions resulted in glare on portions of the mudflat that retained a few centimeters of water. This glare was sufficient to obscure the presence of eelgrass. When doing the regression analyses, virtual ground control plots for 28 August that had obvious glare were deleted. However, the resulting model (Figure 7) was applied to all four dates, including the 28 August imagery, and this makes the prediction of vegetation cover for this date somewhat questionable. Although data from our virtual ground control plots indicate a decline in eelgrass cover for this date (Table 4), it is not clear if our vegetation cover layer for this date (Figure 8d) accurately quantifies the extent of this decline (Table 6), or if it may be compromised by glare in those areas that retained some water. Finally, imagery obtained at the lowest possible tide stage, when eelgrass is exposed rather than submerged, is critical for going beyond simply mapping the presence or absence of eelgrass and quantifying percent cover and distinguishing between different species as was done by Bergner et al. (2025). The ideal set of conditions for obtaining imagery will rarely occur, and therefore, when imagery is obtained under suboptimal conditions, it is critically important to be mindful of how these conditions may affect results.

In addition to evaluating recreational impacts, the approach developed here can also be useful for monitoring the impacts of commercial shellfish operations as well as contributing to a better understanding of seasonal and interannual variation in eelgrass in response to other natural and anthropogenic factors.

Funding for this work was provided by the Whatcom County Marine Resources Committee. Larabee State Park also supported this work. The authors are grateful to Matthew Tatman at Larrabee State Park for providing us with the vehicle access data for Wildcat Cove. The authors also thank Andy Bunn for assistance with the R code and Stefan Freelan for providing GIS support.