Unoccupied aerial vehicle-assisted monitoring of benthic vegetation in the coastal zone enhances the quality of ecological data

Abstract

Traditional monitoring (e.g., in-water based surveys) of eelgrass meadows and perennial macroalgae in coastal areas is time and labor intensive, requires extensive equipment, and the collected data has a low temporal resolution. Further, divers and Remotely Operated Vehicles (ROVs) have a low spatial extent that cover small fractions of full systems. The inherent heterogeneity of eelgrass meadows and macroalgae assemblages in these coastal systems makes interpolation and extrapolation of observations complicated and, as such, methods to collect data on larger spatial scales whilst retaining high spatial resolution is required to guide management. Recently, the utilization of Unoccupied Aerial Vehicles (UAVs) has gained popularity in ecological sciences due to their ability to rapidly collect large amounts of area-based and georeferenced data, making it possible to monitor the spatial extent and status of SAV communities with limited equipment requirements compared to ROVs or diver surveys. This paper is focused on the increased value provided by UAV-based, data collection (visual/Red Green Blue imagery) and Object Based Image Analysis for gaining an improved understanding of eelgrass recovery. It is demonstrated that delineation and classification of two species of SAV (Fucus vesiculosus and Zostera marina) is possible; with an error matrix indicating 86–92% accuracy. Classified maps also highlighted the increasing biomass and areal coverage of F. vesiculosus as a potential stressor to eelgrass meadows. Further, authors derive a statistically significant conversion of percentage cover to biomass (R² = 0.96 for Fucus vesiculosus, R² = 0.89 for Zostera marina total biomass, and R² = 0.94 for AGB alone, p < 0.001). Results here provide an example of mapping cover and biomass of SAV and provide a tool to undertake spatio-temporal analyses to enhance the understanding of eelgrass ecosystem dynamics.

Keywords

UAV remote sensing marine biology coastal monitoring seagrass macroalgae

Introduction

Many studies have shown that seagrass habitats are declining worldwide due to anthropogenic disturbances such as eutrophication, habitat disruption and fragmentation, water turbidity, pollutants, and coastal development (e.g., Flindt et al., 1999; Haynes et al., 2000; Moksnes et al., 2018; Orth et al., 2006; Short et al., 2011; Waycott et al., 2009; Williams, 2007). Many coastal systems suffer from high nutrient loadings coming from agricultural runoff and sewage water, which hamper natural recovery of seagrasses due to growth and shading from epiphytes. In other systems, former seagrass habitats are now occupied by macroalgae, hindering their recovery by shading and ballistic impact from drifting macroalgae (Canal-Vergés et al., 2014; Rasmussen et al., 2012; Valdemarsen et al., 2010, 2011). Within Danish waters, 40 years of historical eutrophication has induced excess sediment organic content such that the sediment anchoring capacity (the ability of the sediment to hold seagrass) also has been lost (Canal-Vergés et al., 2016; Flindt et al., 2016; Kuusemäe et al., 2016; Valdemarsen et al., 2014). Seagrasses are introduced as a “biological quality element in the European Union Water Framework Directive (WFD) for coastal and transitional waters” (WFD, CIS Guidance Document No. 20, 2009) due to the numerous ecosystem services they provide. These include, for example, coastal protection and stabilization of the sediment (Ward et al., 1984), nutrient immobilization (Flindt et al., 1999), habitat formation and primary production (Boström and Bonsdorff, 2000), and carbon sequestration (Hemminga and Duarte, 2000). The objectives for the status of seagrasses in European and transitional waters are described in the WFD, annex V; “changes in angiosperm abundance due to anthropogenic activities should not be detectable” and “the levels of angiosperm abundance should be consistent with undisturbed conditions.” (European Parliament, 2000)

In Denmark, the depth limit of the seagrass species Zostera marina (eelgrass) is implemented as one of three biological quality elements and it is routinely measured (bi-yearly) along limited, 2-m-wide transects located in estuaries and along the open coast (Bruhn et al., 2017). These efforts produce information about the growth conditions related to the light climate at the specific location of sampling; however, they do not provide spatial information about eelgrass coverage and production (e.g., the transects only represents permilles of the coastal system with no information of bed extent). Spatial information about eelgrass coverage is essential to understand how eelgrass recovery occurs in ecosystems that are highly altered when compared to the reference state (Benson et al., 2013; Orth et al., 2012). Only by including knowledge of full system coverage and area-specific production is it possible to evaluate the state of ecosystem services provided by recovering eelgrass meadows. The complex and heavily heterogeneous spatial landscape structures in these systems, including macroalgal assemblages, eelgrass patches, and their varying coverage all influence the ecosystem services provided by benthic vegetation. As such, collecting data with traditional techniques based on observations, photos and video recorded along narrow transects is time and labor intensive and requires considerable effort. The activity fulfills the aim of monitoring the depth limit of eelgrass, but it is not possible to use these observations to represent coverages of eelgrass at a landscape level, as the heterogeneity of these ecosystems hinders inter- and extrapolation between the transects.

Due to historical reduction in eelgrass distribution and biomass, a higher proportion of nutrients is available for macroalgae production in Odense Fjord (Valdemarsen et al., 2010). The macroalgae stocks, especially the highly mobile bladderwrack, Fucus vesiculosus, impact eelgrass beds by a range of stress-effects including deteriorated light environment, erosion, and resuspension (Canal-Verges et al., 2010, Flindt et al., 2007). Mobile macroalgae also has a destrucive ballistic impact on eelgrass seedlings and mature beds when they are dragged by water currents while being attached to bigger stones (Canal-Vergés et al., 2014). Here, F. vesiculosus physically damages seedlings as evidenced by high seedling mortality during periods of intense macroalgae drift (Valdemarsen et al., 2010). Further, when F. vesiculosus enters mature eelgrass patches, it can create additional fragmentation along the patch edges and internal bare areas as quantified by field observations (Flindt et al. 1997, 1999). Given the complex interactions between F. vesiculosus and eelgrasses, tools are required to rapidly and accurately assess the spatial extent and biomass of each SAV to inform management of areas where the two species co-occur.

Studies have shown that traditional satellite- and plane-based remote sensing of SAV has its limitations (Hossain et al., 2015) and that no methodology is suitable for assessing all relevant ecological parameters: presence/absence, cover, species, biomass, and thus change. Due to the variable flight height, and hence resulting image resolution, of UAVs and the general Very High-Resolution (VHR) of UAV-imagery, it is argued that this platform is most suitable for covering more or all these parameters. Especially if the aerial imagery is combined with field measurements of SAV biomass, a complete overview of the ecological state and development can be obtained. In recent years, UAV applications in marine areas have been increasingly utilized: Barrell and Grant (2015) presented a scheme for extracting information of eelgrass and blue mussels from imagery collected using a balloon. Duffy et al. (2018) presented a scheme for analysis of a seagrass and surrounding macroalgae in the intertidal zone. Svane et al. (2016) presented a method for UAV-based mapping of eelgrass and macroalgae in the subtidal zone, and Nahirnick et al. (2019) outlined the benefits and problems in UAV-based mapping of seagrass in coastal zones. With the rapid development of UAV technology and associated software to support image interpretations by segmentation and classification of SAV based on spectra, shape, and context [in short Object Based Image Analyses (OBIA)] (Blaschke, 2010), it is logical for ecological managers to attempt to integrate these new technologies with the data collected from more traditional methods (e.g., video sleigh or diver surveys). UAV-based, georeferenced, image-acquisition covering large areas, makes it possible to perform classifications across large spatial scales in comparatively small timescales (Joyce et al., 2017). By collecting images from high altitudes, which allows imagery collected to cover larger areas at the cost of a reduction in resolution, eelgrass patch structures, and sizes may be characterized. Further, by using UAV-based imagery from low altitudes, UAV systems provide the possibility to assess the density of a patch; which in turn can be used as an indicator of local stressors to eelgrass. Duffy et al. (2018) also states the possibility of complementing spatial information collected (e.g., seagrass and macroalgae coverage) with biomass information to support the growing interest in carbon sequestration studies (e.g., McLeod et al., 2011; Howard et al., 2017; Sousa et al., 2019). This focus point also has relevance in relation to determining the plant bound nutrient transport which impacts estuarine areas (Salomonsen et al., 1999; Flindt et al. 2004, 2007).

The aim of this study was therefore to develop and test UAV-based analytical methods to (1) quantify eelgrass and macroalgae coverage, (2) obtain area-specific biomass estimates of eelgrass and F. vesiculosus, (3) monitor and quantify vegetative growth of eelgrass patches over time, and (4) utilize fragmentation metrics as an indicator of the degree of stress present.

Methods

Study area

Odense Fjord is a shallow estuary with an average depth of 2.2 m impacted by freshwater inputs from River Odense. Depending on freshwater input and exchange with Kattegat Sound, salinity varies from 5 to 25 ppm (Petersen et al., 2009). Odense Fjord has a large catchment area (1046 km²) resulting in substantial nutrient loading primarily due to agricultural runoff. Prior to 1990 the nutrient loading was relatively high (2500 and 300 t N and P y⁻¹), but after the implementation of several water action plans the nutrient loading was reduced to present levels (1500–2000 t N y⁻¹ and 50 t P y⁻¹) (Petersen et al., 2009). Reductions in nutrient loadings have improved the overall water quality, as evidenced by a reduction of phytoplankton concentrations and reduced blooms of opportunistic macroalgae (Petersen et al., 2009). However, Odense Fjord does not meet the targets set out by the WFD regarding eelgrass (Zostera marina) depth limits due to the current nutrient levels. In the reference condition, based on coverage data from 1880–1930, eelgrass covered substantial areas (>15.1 km²) of Odense Fjord at Enebærodde (Petersen et al., 2009), while the coverage today is below 1.2 km². The environmental depth target in the EU WFD Water Management Plan is 4.2 m in Odense Fjord, corresponding to 75% eelgrass coverage in the estuary in the reference state. The eelgrass population has not shown signs of recovery—either in shallow or deeper areas of Odense Fjord—suggesting that light availability is only one of the many pressures affecting the current eelgrass distribution in the system. Valdemarsen et al. (2010) demonstrated that physical stress from water currents and wave action together with ballistic impact from F. vesiculosus drifting along the seabed attached to stones hinder the survival of eelgrass seedlings and hence the natural recovery of eelgrass.

For this study two focus areas in the Northern Part of Odense Fjord were chosen for SAV-mapping and analysis (Figure 1): Site (A) a nursery area for F. vesiculosus in the relatively protected North-West corner of Odense Fjord with a large monospecific standing stock of F. vesiculosus (approximately 400 ha). Site (B) contains a medium sized (0.6 ha) eelgrass population with a high degree of fragmentation and a high-level of physical impact (Canal–Vergés et al., 2014; Valdemarsen et al., 2010). This region is situated in the Northern part of Odense Fjord with a water depth of 0.5–2 m and is heavily exposed to wind and wave forces resulting from prevailing South Westerly winds (Danish Meteorological Institute , 2016). From site (A), large amounts of F. vesiculosus drift to site (B) where it is thought to hinder eelgrass recovery processes, with eelgrass patches continuously fragmented and never able to close the gaps and grow together.

Figure 1.

Map of Denmark (left) showing Odense Fjord and the sites for mapping F. vesiculosus aggregations (A) and eelgrass beds (B) (right). Contains data from the Agency of Data Distribution and Efficiency, Denmark, 2021.

Technical setup

The UAV platforms used in this project were consumer grade quadcopters with no sensor upgrades. The quadcopter type was chosen to be able to fly in alternating altitudes and perform UAV-based ground truthing but still being able to cover large areas. The UAV-model used for mapping F. vesiculosus at site B was a DJI Phantom 3 Professional (P3) with the standard 1/2.3 inch CMOS sensor with 12.4 million effective pixels and a field of view of 94° at a 20 mm f. 2.8 aperture (DJI, 2020). For all mapping at site A, a DJI Phantom 4 Professional (P4) with the standard 1 inch CMOS sensor with 20 million effective pixels and a field of view of 84° at an 8.8 mm/24 mm f. 2.8–11 aperture was used (DJI, 2020). These platforms were chosen as they are both accessible, cost-effective, and durable platforms with a high-quality camera. High capacity 4S LiPo (lithium-polymer) batteries of 5350–5870 mAh were used providing a practical flight time of 22–28 min depending on the wind conditions. For security reasons, no battery was used below the 20% mark before return-to-land. This made mapping up to 15 ha per battery possible when flying in 100 m altitude.

Field procedure

Biomass study: To determine the relationship between the top-down surface area visible from aerial images and the biomass of F. vesiculosus, 15 samples of complete specimens were collected at site A (593,725.58 m E; 6,153,864.01 m N) in July 2015 (Figure 1). The samples were collected randomly at different water depths (0.4–1.5 m) with samples exhibiting a variety of underwater shapes. The algal material was collected as single individuals and brought back to the laboratory. Here the specimens were carefully cleaned with demineralized water and epifauna, epiphytes, and debris were removed. They were then oven-dried at 60°C until constant weight. For each sample, the wet weight and dry weight were measured.

To determine the relationship between coverage, shoot number and Above Ground Biomass (AGB) of eelgrass, samples were collected at 16 randomly chosen plots at site B (Figure 1) in September 2016. Sampling involved placing a 25 x 25 cm metal frame in the chosen plots and collecting all eelgrass biomass inside the frame. Plant material were then carefully cleaned and processed as described for macroalgae above. The dry weight of the material was then determined and calculated per square meter. The POC (particulate organic carbon), C:N ratio and thus nutrient content was calculated from conversions outlined in the published literature (Duarte, 1990; Kristensen, 1990). At both sites, the F. vesiculosus specimens and eelgrass sampling sites were overflown at 20- and 100-m altitude in connection with the biomass sampling to be able to couple these. For all 100-m flights, it was assumed that eelgrass exhibited a percentage cover of 80% to standardize biomass calculations and prevent overestimation.

UAV flight: Based on preliminary test-flights a general mapping altitude of 100 m corresponding to a Ground Sample Distance (GSD) of 3.8 cm (P3) and 2.6 cm (P4) was chosen. This provided a sufficient resolution for general classification on a landscape level. For studying classification efficiency of different coverages of eelgrass, flights in 20 m altitude were performed. In each area, UAV-based ground truthing was tested and subsequently performed at 2–10 m altitude (GSD: 0.05–0.38 cm) which was used for classification, validation, and studies of classification limitations.

Image analysis

The production of georeferenced orthomosaics from single images was performed in Agisoft Photoscan ver. 1.3.4 and Agisoft Metashape ver. 1.5.5 (Agisoft, 2018; Agisoft, 2020). This photogrammetry software presents an easy-to-use workflow for stitching single images through steps covering correction, point cloud generation, surface construction, and orthomosaic production. All segmentation and classification of image objects was performed using the software Trimble eCognition Developer ver. 9.3 and 9.5 (Trimble Inc, 2020) by an OBIA approach (Blaschke, 2010; Lang, 2008) in which spectral information is combined with shape, distances, and context of pixels clustered into objects. All parameters in this analysis were determined by preceding testing to establish the optimal range for segmentation and recognition of the species of interest. For initial segmentation, a multiresolution algorithm was used (Baatz and Schape, 2000) with the adjustable parameters: scale (100 m imagery = 70, 20 m imagery = 50), color/shape ratio (0.2), compactness (0.5) and equal weight of each color band (1,1,1). To secure an objective and comparable classification for different time-steps and between stations, a supervised machine learning classification approach with training samples was used (Lillesand et al., 2015). A Support Vector Machine (SVM) algorithm was chosen for the machine learning approach as it is precise (Tzotsos and Argialas, 2008) and less sensitive to low numbers of training samples than other machine learning algorithms (Myburgh and van Niekerk, 2014). In all analyses, a number of 20 samples per class was assessed as the optimum and applied for training. The object features used for classification were the mean and standard deviation of the spectral RGB (Red, Green, Blue) values of the pixels aggregated within each object and the mean difference to neighboring objects (distance = 0). The following classes were used in the classification: Macroalgae (here: F. vesiculosus) “Fv," Eelgrass "‘Zm," Bare bottom “BBx” and the specific eelgrass coverage classes for low altitude mapping: Coverage of 1–50% “Zm_SUB_50," between 50 and 75% “Zm_50_75” and above 75% “Zm_75_100." Several classes of bare bottom were used and subsequently merged due to variance in their spectral features. A class covering solar glints was also introduced but did not have an impact on the analysis as this was performed on orthomosaics in which glints were removed.

For classification of F. vesiculosus at site (A) high altitude (100 m) orthophotos from September 2015 were used. For classification of eelgrass and F. vesiculosus at site (B), high altitude (100 m) orthophotos created from imagery acquired in the growth season 2015–2017 were used, and for detailed eelgrass coverage percentage classification at site (B) low altitude (20 m) orthomosaics created from imagery acquired in the growth season of 2016 were used.

Accuracy assessment: To assess the accuracy of classifications, a confusion matrix (Congalton and Green, 2009; Foody, 2008) was constructed using the built-in accuracy assessment module in eCognition. Following the initial classification, 10 subsets for each classification type were chosen in a stratified random manner for Fv (100 m) at site (A), Zm and Fv (100 m) at site (B), and low altitude Zm, Fv (20 m) at site (B). The classification accuracy was then determined from 100 (in 100 m imagery) and 50 (in 20 m imagery) validation points for each class chosen by stratified random sampling based on very-low altitude imagery (<2 m).

Statistical procedures

Relationships between UAV-based observations and biomasses of both F. vesiculosus and eelgrass together with eelgrass shoot density, was determined with Pearson’s linear correlation analysis. Data was tested for normal distribution (Shapiro-Wilk’s test) prior to the correlation analysis. All statistical procedures were performed in SigmaPlot v. 12 (SSI, 2019) using a significance level p < 0.05.

Results

Results are divided into delineation and classification—including accuracy assessment—of the monitored organisms and objects of interest, biomass calculations of F. vesiculosus and eelgrass along with a combination of the two sections.

Delineation and classification

Both eelgrass and F. vesiculosus were delineated and classified in this study using the described methods. In the F. vesiculosus-dominated site (A), it was possible to extract the coverage (as summed top-down surface area) of the F. vesiculosus (Fv) along with separating individual and accumulated macroalgae from the bare bottom (BB) (Figure 2).

Figure 2.

Left: Unclassified map of site (A) on landscape (100 m) and species (20 m) levels, respectively. Right: The same area after delineation and classification into three distinct classes, F. vesiculosus (Fv) and Bare Bottom (BB1 and BB2).

For the F. vesiculosus study at site (A) three classes were used: macroalgae denoted “Fv” (F. vesiculosus), bare bottom 1 (BB1) and bare bottom 2 (BB2). As the sediment surface had varying spectral properties in the project area, due to varying organic content and/or coverage of benthic diatoms, two sediment-classes were needed. Using this delineation, it was possible to determine the coverage and total area of F. vesiculosus. Further, when combined with dry weight measurements (outlined below), it was possible to predict the biomass of F. vesiculosus in the project area (Table 1).

Table 1.

Result of the confusion matrices made for all sub-studies (Site (A) and (B) at 100 m and 20 m altitude) and objects of interest (eelgrass—Zm, F. vesiculosus—Fv, and bare bottom—BB) (n = 10). The classification accuracy is expressed as Kappa Indices of Agreement (KIA) and overall accuracy (O.A), User and Producer values and calculated standard errors.

Site (A) 100 m			Site (B) 100 m
Class	Fv	BB	Class	Zm	Fv	BB
Producer	0.92 ± 0.03	0.99 ± 0.01	Producer	0.91 ± 0.05	0.83 ± 0.04	0.99 ± 0.01
User	0.97 ± 0.01	0.92 ± 0.02	User	0.95 ± 0.04	0.85 ± 0.03	0.95 ± 0.02
KIA	92 ± 3%		KIA	92 ± 3%
Overall	94 ± 1%		Overall	93 ± 4%
Site (B) 20 m
Class	Zm_SUB_50	Zm_50_75	Zm_75_100	BB	Fv
Producer	0.81 ± 0.03	0.86 ± 0.02	0.87 ± 0.03	0.92 ± 0.03	0.68 ± 0.03
User	0.65 ± 0.08	0.83 ± 0.03	0.96 ± 0.01	0.91 ± 0.03	0.87 ± 0.04
KIA	83 ± 3%
Overall	86 ± 1%

For the monitored eelgrass beds at site (B), it was possible to delineate and classify objects of interest at varying levels of detail depending on flight altitude. For this study area, low altitude (10 m flight height) a total of 7 classes were used. It was possible to distinguish between different coverage classes and separate eelgrass from both bare bottom and F. vesiculosus (Figure 3).

Figure 3.

Left: Unclassified image of two different eelgrass beds captured at 20 m flight height at site (B). Right: The same images after delineation and classification into seven distinct classes consisting of varying eelgrass coverages (EG_SUB_50 [< 50 % cover], EG_50_75 [50–75 % cover], EG_75_100 [75–100 % cover], F. vesiculosus [Fv], Bare Bottom 1 and 2 and Reflection.

All patch edges are classified as 1–50% coverage as expected and even small variations in degree coverage were detected.

For determining eelgrass and F. vesiculosus general coverage and biomass on a large scale using high altitude images obtained at 100 m, four classes were used: Eelgrass (Zm), F. vesiculosus (Fv) and two classes of bare bottom (BB1 and BB2) (Figure 4).

Figure 4.

Top row: Unclassified images of eelgrass beds at site (B) from 2015, 2016, and 2017 obtained in 100 m flight height. Bottom row: The same images after delineation and classification into four distinct classes: Eelgrass (Zm), F. vesiculosus (Fv), and bare bottom (BB1 and BB2).

Analyses showed a strong ability to delineate between eelgrass and bare bottom, including even macroalgae at the edges of the eelgrass beds. In zones with low coverage of eelgrass or macroalgae, misclassifications occurred (Figure 4).

Accuracy assessment

For all analyses performed on the varying flight heights and species an accuracy assessment was made (Table 1). An n = 10 and 50–100 validation samples per class were used for all classification types to secure a valid foundation for the assessment.

The accuracy for classification of high-altitude imagery at site (A) and (B) was 94 ± 1% and 93 ± 4% with KIA (Kappa Indices of Agreement) values of 92 ± 3% and 92 ± 3%, respectively, the latter expressing the probability of agreement with expected values (Congalton and Green, 2009). The low altitude imagery (site (B) 20 m) was >86 ± 1% expressed as overall accuracy and 83 ± 3% expressed as KIA. At site (A), the largest source of error was F. vesiculosus being omitted (producer accuracy of 0.92 ± 0.03) and bare bottom being committed (user accuracy of 0.92 ± 0.02). For the high-altitude imagery at site (B), the largest source of error was F. vesiculosus being omitted (producer accuracy of 0.83 ± 0.04) instead being classified as eelgrass. For the low-altitude imagery at site (B), the largest source of error was eelgrass with a 1–50% coverage percentage being committed as either F. vesiculosus or eelgrass in with a higher coverage percentage (user accuracy of 0.65 ± 0.08). F. vesiculosus was equally being omitted (producer accuracy of 0.68 ± 0.03) due to being misclassified as low-coverage eelgrass or bare bottom.

Biomass study

It was possible to determine a direct and significant correlation between the surface area of bladderwrack (Fv) and its biomass in gDW (Figure 5) with a p-value < 0.001 an R²-value of 0.96.

Figure 5.

F. vesiculosus (Fv) biomass (y-axis) in relation to visible top-down surface area (x-axis).

The correlation shows low variation between the individual F. vesiculosus specimens, with small individuals ranging 16–51 gDW to large individuals of 436–619 gDW biomass with surface areas ranging from 0.02 to 0.37 m².

It was equally possible to determine significant correlations for eelgrass. Significant correlations between percentage cover and shoot number along with degree of coverage and total biomass (AGB and BGB) in gDWm⁻² were shown with p-values of <0.001 and R²-values of 0.95 and 0.89, respectively (Figure 6).

Figure 6.

Left: Eelgrass [Z. marina (Zm)] degree of coverage in relation to shoot number. Right: Total Zm biomass in relation to percentage cover.

At this specific site a coverage degree of eelgrass of 25% corresponded to a shoot number of 192 m⁻² and 95% coverage to a shoot number of 1072 m⁻². When converting biomass (AGB only) to coverage a range from ∼25 gDWm⁻² at < 10% coverage to ∼259 gDWm⁻² at > 95% coverage could be measured. For Below Ground Biomass (BGB), the numbers ranged from ∼10 gDWm⁻² at < 10% coverage to ∼646 gDWm⁻² at > 95% coverage.

Using the classified images, it was possible to determine the zones where of highest physical impact. Further, once classifications were combined with the dry weight measurements it was then possible to calculate the total biomass of the eelgrass patches in the 2000 m² survey area (Table 2).

Table 2.

Overview of the parameters calculated from the classified images from 2015–2017: Area, Perimeter, Perimeter-to-Area ratio (PA-ratio) of eelgrass (Zm), Coverage of F. vesiculosus (Fv), Total ABG biomass of eelgrass at 80% coverage in kgDW, and Total biomass of F. vesiculosus in kgDW in the 2000 m² survey area.

Year	Area Zm (m²)	Perimeter (m)	P/A-ratio	Area Fv [% of total]	Biomass Zm [kgDW]	Biomass Fv [kgDW]
2015	278	280	1.00	11.2	52.8	377.9
2016	300	279	0.93	5.4	55.8	177.8
2017	334	305	0.91	8.1	63.2	320.7

The analysis shows a clear growth of the eelgrass beds from 2015 to 2017, with the main patch growing from 279 to 334 m². This is an increase in biomass of 10 kg DW (AGB), assuming an eelgrass percentage cover of 80% corresponding to 0.25 kg N and 0.038 kg P (Duarte, 1990). The P/A-ratio (perimeter/area) decreased nearly 10% during the 3 years indicating decreasing complexity of the seagrass bed outer border. Regarding the macroalgae in the focus area, the analysis provides a snapshot of the specific coverage and thus accumulated biomass, which varies from 177 kg DW in 2016 to above 320 kg DW in 2015 and 2017, corresponding to 3.84–6.94 kg N and 0.34–0.62 kg P when reference nutrient values are coupled to the coverages (Kristensen, 1990; Lorenzo et al., 2017).

Discussion

General findings: The study illustrated the feasibility of using Very High-Resolution (VHR) imagery for near coast mapping of submersed aquatic vegetation on both a landscape level at a high altitude and for ground truthing of individual species at lower altitudes. It was possible to achieve positive results for all the part-aims of this study. Delineation and thus determination of coverage of eelgrass and bladderwrack was achievable using the proposed methodologies and eCognition software for analysis. Compared to earlier studies (Barrell and Grant, 2015; Duffy et al., 2018) our study was performed in a subtidal area with constant water coverage which could pose issues (reflection, turbidity etc.) as described by Nahirnick et al. (2019). However, our study showed that the influences of these issues can be reduced by using a multi-feature approach made possible by using the OBIA-framework in which more reliance on shape and context is introduced. This finding is consistent with other literature, for example, (Tzotsos and Argialas, 2008) who showed a high accuracy of 86–92% and generally demonstrated robustness towards changing environmental conditions when mapping using an OBIA approach. The largest source of error was separating 1–50% coverage percentage eelgrass from macroalgae and bare bottom in low-altitude imagery (20 m) but for generally separating eelgrass, bare bottom and F. vesiculosus in high-altitude imagery (100 m) high accuracies (<90%) was found.

Ecosystem services: Delineation and classification processes provided distinct classes consisting of bare bottom, macroalgae, and several classes of varying eelgrass coverages making analysis of recovery and stress possible (Figure 5). Collection of spatial data on eelgrass and macroalgae coverage across space and time is essential to understand how eelgrass is recovering in ecosystems that are highly altered when compared to the ecosystems where eelgrass is the dominating benthic primary producer (Benson et al., 2013, Flindt et al., 2016, Kuusemäe et al., 2016; Orth et al., 2012). Only by including knowledge of full system coverage and area-specific production is it possible to evaluate the ecosystem services that eelgrass is providing with respect to self-protection, immobilizing nutrients during the growth season and improving the light climate by reducing resuspension events. In Figure 4, delineation and classification of the images verify the dynamic of drifting macroalgae in an eelgrass area within 2015–2017, where the area-specific macroalgae density and geographic position of individuals varies between years. Inflowing F. vesiculosus is most likely responsible for the variable growth dynamics (e.g., increasing P/A-ratio) of the eelgrass beds here due to its inherent stress-effect (Canal-Verges et al., 2010; 2013; Flindt et al., 2007) and the inability of new patches to develop from seedlings in the inner part of the system (Valdemarsen et al., 2010).

Regarding the composition of the eelgrass patches in the area, highly detailed delineation is important as this provides information about ecosystem services provided by the patches along with information of stress-related processes. Highly fragmented patches of low size and high abundance, for example, have value as microhabitats (Hirst and Attrill, 2008) but are also an indicator of high physical stress (Fonseca and Bell, 1998). As edge effects and gap dynamics play a role in supporting faunal biodiversity (especially in relation to arthropods) (Tanner, 2005; Warry et al., 2009) and overall community structure (Bell et al., 1999), determining these is important in understanding ecosystem functioning. Using the methods outlined in this study, it will be possible to monitor the mortality of small, newly established patches, which is known to be high (Olesen and Sand-Jensen, 1994). From 2015 to 2017, an areal expansion of 20.1% (7.9–11.3% y⁻¹) was observed in the monitored area (Table 2). This rate is comparable to existing studies showing 19% y⁻¹ in larger patches while smaller patches show a much higher growth rate (Olesen and Sand-Jensen, 1994). The lower yearly growth in the study area corresponds well with the higher exposure to physical stress in this site.

Having the ability to precisely estimate the standing stock of eelgrass and macroalgae biomass and thus nutrient content is important both in defining the value of ecosystem services and monitoring seagrass restoration success (Lange et al., in prep; Orth et al., 2020); to calculate the effectiveness of these habitats to sequester blue carbon (Howard et al., 2017; Mazarrasa et al., 2018; Sousa et al., 2019), and for validating ecological models (Kuusemäe et al., 2016). Macroalgae also have a somewhat undisclosed role as carbon sinks and as such precise biomass mapping may support future studies aiming to quantify this role (Krause-Jensen and Duarte, 2016).

Seagrasses are essential to achieve stable and well-working shallow estuarine and coastal ecosystems due to their numerous ecosystem functions and services (Flindt et al., 2016; Orth et al., 2012). Most of these services are associated to growth related processes, where increasing shoot, rhizome and root densities in existing patches, but also expansion into bare bottom areas creates more seagrass biomass. Increases in biomass immobilizes equivalent masses of dissolved inorganic nutrients (DIN and DIP) taken up by the plant for growth purposes. By combining UAV and field sampling, it was possible to achieve highly statistically significant correlations between surface areas of F. vesiculosus (R² = 0.96 and p < 0.001) and eelgrass (R² = 0.94 and 0.89, p < 0.001) and area-specific biomasses. For F. vesiculosus, the correlation was seemingly unaffected by the varying underwater shape of the specimens. The measured biomass of F. vesiculosus was between 600 and 700 gDW for specimens of app. 0.4 m² in size which correspond well with known values of 800–2000 gDW m⁻² at 80–100% coverage (Öberg, 2006).

For eelgrass, the measured AGB of 250 gDW m⁻² and total biomass of 900 gDW m⁻² at 100% coverage (app. shoot density of 1300 m⁻²) corresponds well with other studies (200-400 gDW m⁻²) and total biomass of 700–800 gDW m^-2 (Olesen, 1999; Olesen and Sand-Jensen, 1994). As we are focusing on one defined area with constant sediment conditions, the biomass shown is the total of both AGB and BGB as the relationship between the two biomass types is expected to be stable as well. The composition of the sediment, especially regarding organic content, plays a role in not only anchoring capacity but also eelgrass rhizome growth (Wicks et al., 2009), thus for studies in different areas unique relationships for ABG and BGB should be made. At Enebærodde (site (B), Figure 1) the organic content is relatively low (0.5–0.9% LOI) (Valdemarsen et al., 2010). Using known values of POC, C:N and P (Duarte, 1990; Kristensen, 1990; Lorenzo et al., 2017), both the plant bound nutrients and the potential as nutrient sinks can be calculated for the two species.

These findings support using image derived estimation of biomass, nutrient content, and temporal change for future studies along with calculations of ecosystem services where improved eelgrass coverage supports less resuspension and a higher benthic light intensity (Orth et al., 2012). It is well-known that eelgrass coverage-to-biomass ratios vary depending on leaf length (Carstensen et al., 2016). Populations growing in areas with suboptimal light intensities (higher turbidity, deeper growing populations) usually produce longer leaves. It is therefore recommended to check the coverage-to-biomass ratio between sites to ensure that trends used are applicable to site-specific growth forms.

Imagery for ground truthing: Traditional ground truthing and validation of classification accuracy is based on on-ground sampling and often additional species-specific information (shoot length, percentage cover etc.) is also collected (Duffy et al., 2018; Husson et al., 2016; Nahirnick et al., 2019). When identifying and validating species presence we argue that very-low altitude, high-resolution UAV-based images are sufficient as a method of ground truthing, as long as the water depth, turbidity, and wave action is not disruptive of a clear image. We found that this approach could be used in approximately 90% of the survey situations making it a fast and easy alternative to manual ground truthing. However, this is an active area of discussion and care should be taken as described by Congalton and Green (2009). Alternatively, drone-based underwater ground truthing can be performed by either lowering a camera underneath the surface (Stæhr et al., 2019) or landing on the surface (Svane et al., 2018). A common trait of traditional ground truthing methods is their relatively high time demand compared to in-air ground truthing.

Implementation: Regarding implementation of the proposed methodology two steps are time consuming and need attention by the user: Collection of samples for classification based on ground truthing images, and collection of samples for validation. These steps are possible to automate in a higher degree by automatically labeling the class in ground truth images and creating a training point from the image meta data for location, but this is a work in progress. Generally, the rulesets used are universally applicable both in time and space, as long as training and validation are repeated for each site/time. The machine learning algorithm is based on spectral and context features which when applied on objects obtain a spatial dimension. Due to an approach where neighboring objects are compared in a certain feature space, the specific environmental conditions on the time of image-acquisition minimally impact the resulting classification. Regarding the timeframe for UAV-based monitoring, a thorough evaluation of Danish weather data from 2013 to 2015 was made in another project (Svane, unpub.). This showed that mapping would be possible in 69–75% of the daytime not taking turbidity into account. Windy conditions also show a limited impact on recognition of eelgrass and F. vesiculosus specifically within Danish coastal waters (Svane, unpub.).

Compared to traditional monitoring methodologies—for example, diver and ROV (Bruhn et al., 2017)—the UAV-based methodology differs in important aspects: The coverage of eelgrass or macroalgae is area-based for the whole study area instead of being based on points in a narrow transect, and the data processing is less time consuming as it can be highly automated. Compared to aerial and satellite-based (DHI, 2019; Frederiksen et al., 2004; Hogrefe et al., 2014) monitoring of SAV the UAV-based method provides a higher spatial resolution in a trade off with spatial extent. The improved spatial resolution (0.5–5 cm/pix) makes more accurate species delineation and classification possible and is thus a prerequisite for precise monitoring efforts. The method is equally suited for training and/or validation for national monitoring projects based on satellite imagery such as the Danish Aquatic Vegetation Map from 2019 (DHI, 2019) or if plane-based aerial imagery is used (Ørberg et al., 2018). Very High-Resolution (VHR) UAV-based imagery with a GSD of 2–3 cm/pixel will in this case support the differentiation between species such as perennial macroalgae and seagrasses along with “dark” zones such as mussel beds or channels.

Due to the ability to distinguish between vegetation types and, at the same time, perform both UAV-based ground truthing services and covering areas of up to 500–800 ha day⁻¹ per UAV, we propose the method as a valid alternative or support to traditional SAV-mapping. As the majority of eelgrass populations inhabit coastal zones with low water levels the total area of interest is also limited, thus the traditional diver and ROV surveys will only be necessary for the deepest growth zones where UAV-camera spectra are unlikely to create distinguishable differences (Hansen, 2018).

For future studies we recommend further development of machine learning rulesets. This includes a shift towards more automated analyses in which neural networks and self-learning will be an important part. For intertidal studies, a multispectral NIR-setup is highly recommended as the additional bands will strengthen classification accuracy (e.g., Barillé et al., 2010), while it has little use in subtidal settings due to the rapid absorption of light in the NIR-spectrum (Cho and Lu, 2010). Preliminary studies (Svane, unpub.) here show a distinct increase in delineation accuracy when using NIR-imagery of eelgrass beds with a maximum of 10–20 cm water coverage, limiting the practical use of the NIR-bands in large scale coastal mapping.

Conclusion

Using UAVs for marine monitoring in the subtidal zone using the proposed methodology is a feasible way of assessing both the classified species/object of interest, percentage cover, density, and biomass of SAV and thus the degree of potential stress from, in this case, drifting macroalgae on vulnerable species such as eelgrass based on common fragmentation parameters (e.g., perimeter-to-area ratio). Methods here may be a useful tool for identifying areas suitable for restoration efforts and, by combining areas with specific biomass samples, it is also possible to assess the overall amount of nutrients and carbon contained in a local ecosystem along with local transport. We therefore propose that this method is implemented on a broader scale in national surveys and monitoring as it heightens the data quality and information level available and can thus make the basis for decision making more qualitative. In the future, specific stress-assessments and estimation of biomass-to-area relationships specific to more SAV-species and sites should be implemented to gain a better understanding of the fine scale landscape processes affecting shallow marine environments. Finally, this new technique seems to have potential to optimize site selection for eelgrass restoration activities and provide a fast and easy-to-use tool for coastal environmental mapping (Lange et al., 2020, unpub).

Footnotes

Acknowledgments

We thank the estate ‘Hofmansgave’ for access to the privately owned project area.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Logistics

Location of the study sites: Site (A): (593,712.77 m E; 6,153,878.78 m N, zone 32N), site (B): (597,283.14 m E; 6,153,819.83 m N, zone 32N). All drone flights were performed in compliance with the current Danish legislation consisting of the Act of Drone flight in built-up areas (The ministry of traffic and buildings, 2016) and the Act of drone flight outside built-up areas (The ministry of traffic and buildings, 2017) by a certified drone pilot. Flights were always performed within VLOS (Visual Line of Sight) with a distance to the UAV never exceeding 600 m.

ORCID iD

Niels Svane

References

Agisoft (2020) Agisoft. Russia: Agisoft LLC. Available at: www.agisoft.com (accessed 01 10 2020).

Agisoft (2018) Agisoft PhotoScan User Manual: Professional Edition. Russia: Agisoft LLC Russia. Available at: http://www.agisoft.com/pdf/photoscan-pro_1_4_en.pdf (accessed 10 March 2021).

Baatz

Schape

(2000) Multiresolution segmentation-An optimization approach for high quality multi-scale image segmentation. In: AGIT Symposium Angewandte Geographische Informationsverarbeitung XII. http://www.ecognition.com/sites/default/files/405_baatz_fp_12.pdf (accessed 18 April 2020).

Barillé

Robin

Harin

, et al. (2010) Increase in seagrass distribution at bourgneuf bay (France) detected by spatial remote sensing. Aquatic Botany 92(3): 185–194. DOI: 10.1016/j.aquabot.2009.11.006.

Barrell

Grant

(2015) High-resolution, low-altitude aerial photography in physical geography: A case study characterizing eelgrass (Zostera marina L.) and blue mussel (Mytilus edulis L.) landscape mosaic structure. Progress in Physical Geography 39(4): 440–459. DOI: 10.1177/0309133315578943.

Bell

Robbins

Jensen

(1999) Gap dynamics in a Seagrass landscape. Ecosystems 2(6): 493–504. DOI: 10.1007/s100219900097.

Benson

Schlezinger

Howes

(2013) Relationship between nitrogen concentration, light, and Zostera marina habitat quality and survival in southeastern Massachusetts estuaries. Journal of Environmental Management 131C: 129–137. DOI: 10.1016/j.jenvman.2013.09.033.

Blaschke

(2010) Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing 65(1): 2–16. DOI: 10.1016/j.isprsjprs.2009.06.004.

Boström

Bonsdorff

(2000) Zoobenthic community establishment and habitat complexity–the importance of seagrass, shoot-density, morphology and physical disturbance for faunal recruitment. Marine Ecology Progress Series 205: 123–138. DOI: 10.3354/meps205123.

10.

Bruhn

Høgslund

Dahl

, et al. (2017) Ålegræs og anden Vegetation på Kystnær Blød Bund (Danish) Technical Report M18. Roskilde: DCE, pp. 1–28.

11.

Canal-Vergés

Petersen

Rasmussen

, et al. (2016) Validating GIS tool to assess eelgrass potential recovery in the Limfjorden (Denmark). Ecological Modelling 338: 135–148. DOI: 10.1016/j.ecolmodel.2016.04.023.

12.

Canal-Vergés

Potthoff

Hansen

, et al. (2014) Eelgrass re-establishment in shallow estuaries is affected by drifting macroalgae–Evaluated by agent-based modeling. Ecological Modelling 272: 116–128. DOI: 10.1016/j.ecolmodel.2013.09.008.

13.

Canal Vergés

Vedel

Valdemarsen

, et al. (2010) Resuspension created by bedload transport of macroalgae: implications for ecosystem functioning. Hydrobiologia. 649: 69–76. DOI: 10.1007/s10750-010-0259-y.

14.

Carstensen

Krause-Jensen

Balsby

TJS

(2016) Biomass-cover relationship for eelgrass meadows. Estuaries and Coasts 39(2): 440–450. DOI: 10.1007/s12237-015-9995-6.

15.

Cho

(2010) A water-depth correction algorithm for submerged vegetation spectra. Remote Sensing Letters 1(1): 29–35. DOI: 10.1080/01431160903246709.

16.

Congalton

Green

(2009) Assessing the Accuracy of Remotely Sensed Data - Principles and Practices (2nd edition), Boca Raton: CRC Press Taylor & Francis Group, 121–127.

17.

Danish Meteorological Institute (DMI) (2016) Extracted Dataset of the Years 2009-2014. Denmark: Danish Meteorological Institute (DMI).

18.

DHI (2019) DHI. Denmark: DHI Grass A/S. Available at: http://satlas.dk/marine-vegetation/ (accessed 01 March 2020).

19.

DJI (2020) DJI. China: SZ DJI Technology Co Ltd. Available at: www.dji.com (accessed 30 October 2020).

20.

Duarte

(1990) Seagrass nutrient content. Marine Ecology Progress Series 67(2): 201–207. DOI: 10.3354/meps067201.

21.

Duffy

Pratt

Anderson

, et al. (2018) Spatial assessment of intertidal seagrass meadows using optical imaging systems and a lightweight drone. Estuarine, Coastal and Shelf Science 200: 169–180. DOI: 10.1016/j.ecss.2017.11.001.

22.

European parliament (2000) Directive 2000/60/ec of the European parliament and of the council of 23 October 2000 establishing a framework for community action in the field of water policy. Official Journal of the European Communities L327/1: 1–71.

23.

Flindt

Salomonsen

Carrer

, et al. (1997) Loss, growth and transport dynamics of Chaetomorpha aerea and Ulva rigida in the Lagoon of Venice during an early summer field campaign. Ecological Modelling 102(1): 133–141. DOI: 10.1016/S0304-3800(97)00093-8.

24.

Flindt

Pardal

MÂ

Lillebø

, et al. (1999) Nutrient cycling and plant dynamics in estuaries: A brief review. Acta Oecologica 20(4): 237–248. DOI: 10.1016/S1146-609X(99)00142-3.

25.

Flindt

Neto

Amos

, et al. (2004) Plant bound nutrient transport. Mass transport in estuaries and lagoons. In: Nielsen SL, Banta GT, Pedersen MF. The influence of primary producers on estuarine nutrient cycling (93-128). Dordrecht: Kluwer Academic Publishers.

26.

Flindt

Pedersen

Amos

, et al. (2007) Transport, sloughing and settling rates of estuarine macrophytes: Mechanisms and ecological implications. Chemical Society Reviews 27(8): 1096–1103. DOI: 10.1016/j.csr.2006.08.011.

27.

Flindt

Rasmussen

Valdemarsen

, et al. (2016) Using a GIS-tool to evaluate potential eelgrass reestablishment in estuaries. Ecological Modelling 338: 122–134. DOI: 10.1016/j.ecolmodel.2016.07.005.

28.

Fonseca

Bell

(1998) The influence of physical setting on seagrass landscapes near Beaufort North Carolina, USA. Marine Ecology Progress Series 171: 109–121. DOI: 10.3354/meps171109.

29.

Foody

(2008) Harshness in image classification accuracy assessment. International Journal of Remote Sensing 29(11): 3137–3158. DOI: 10.1080/01431160701442120.

30.

Frederiksen

Krause-Jensen

Holmer

, et al. (2004) Long-term changes in area distribution of eelgrass (Zostera marina) in Danish coastal waters. Aquatic Botany 78(2): 167–181. DOI: 10.1016/j.aquabot.2003.10.002.

31.

Hansen

(ed.), (2018) Marine Områder 2016 NOVANA (Danish). Aarhus Universitet, DCE–Nationalt Center for Miljø og Energi–Scientific Report, pp. 1–140. Available at: http://dce2.au.dk/pub/SR253.pdf (accessed 05 April 2020).

32.

Haynes

Müller

Carter

(2000) Pesticide and herbicide residues in sediments and seagrasses from the great barrier reef world heritage area and queensland coast. Marine Pollution Bulletin 41(7–12): 279–287. DOI: 10.1016/S0025-326X(00)00097-7.

33.

Hemminga

Duarte

(2000) Seagrass Ecology. UK: Cambridge University Press.

34.

Hirst

Attrill

(2008) Small is beautiful: An inverted view of habitat fragmentation in seagrass beds. Estuarine Coastal and Shelf Science 78(4): 811–818. DOI: 10.1016/j.ecss.2008.02.020.

35.

Hogrefe

Ward

Donnelly

, et al. (2014) Establishing a baseline for regional scale monitoring of eelgrass (Zostera marina) habitat on the lower alaska peninsula. Remote Sensing 6(12): 12447–12477. DOI: 10.3390/rs61212447.

36.

Hossain

Bujang

Zakaria

, et al. (2015) The application of remote sensing to seagrass ecosystems: an overview and future research prospects. International Journal of Remote Sensing 36(1): 61–113. DOI: 10.1080/01431161.2014.990649.

37.

Howard

McLeod

Thomas

, et al. (2017) The potential to integrate blue carbon into MPA design and management. Aquatic Conservation: Marine and Freshwater Ecosystems 27(S1): 100–115. DOI: 10.1002/aqc.2809.

38.

Husson

Ecke

Reese

(2016) Comparison of manual mapping and automated object-based image analysis of non-submerged aquatic vegetation from very-high-resolution UAS images. Remote Sensing 8(9): 1–18. DOI: 10.3390/rs8090724.

39.

Joyce

Duce

Leahy

, et al. (2017) Principles and practice of acquiring drone-based image data in marine environments. Marine and Freshwater Research 70(7): 952–963. DOI: 10.1071/MF17380.

40.

Krause-Jensen

Duarte

(2016) Substantial role of macroalgae in marine carbon sequestration. Nature Genetics 9: 737–742.

41.

Kristensen

(1990) Characterization of biogenic organic matter by stepwise thermogravimetry (STG). Biochemistry 9(2): 135–159. DOI: 10.1007/BF00692169.

, et al. (2016) Modelling stressors on the eelgrass recovery process in two Danish estuaries. Ecological Modelling 333: 11–42. DOI: 10.1016/j.ecolmodel.2016.04.008.

43.

Lang

(2008) Object-based image analysis for remote sensing applications: modeling reality–dealing with complexity. In: Blaschke

Lang

Hay

(eds.), Object-Based Image Analysis – Spatial Concepts for Knowledge-Driven Remote Sensing Applications. Berlin Heidelberg, Germany: Springer-Verlag, pp. 3–27.

, et al. Large-scale eelgrass transplantation: A measure for carbon and nutrient sequestration in estuaries. Unpub. accepted 09-2021 in Marine Ecology Progress Series: 1-12.

, et al. (2020) Protection against physical stress enhances eelgrass (Zostera marina) transplant performance. In Development and Test of Measures Supporting Recovery of Eelgrass (Zostera marina) in Estuaries - Phd Thesis, Lange T (2020). Odense, Denmark, University of Southern Denmark, 60-74.

(2015) Remote Sensing and Image Interpretation. USA: John Wiley & Sons, 485–608.

, et al. (2017) Proximate composition and nutritional value of three macroalgae: ascophyllum nodosum, fucus vesiculosus and bifurcaria bifurcata. Marine Drugs 15(11): 1–11. DOI: 10.3390/md15110360.

, et al. (2018) Habitat characteristics provide insights of carbon storage in seagrass meadows. Marine Pollution Bulletin 134: 106–117. DOI: 10.1016/j.marpolbul.2018.01.059.

, et al. (2011) A blueprint for blue carbon: Toward an improved understanding of the role of vegetated coastal habitats in sequestering CO2. Frontiers in Ecology and the Environment 9(10): 552–560. DOI: 10.1890/110004.

, et al. (2018) Local regime shifts prevent natural recovery and restoration of lost eelgrass beds along the Swedish West Coast. Estuaries and Coasts Estuaries and Coasts 41(6): 1712–1731. DOI: 10.1007/s12237-018-0382-y.

51.

Myburgh

van Niekerk

(2014) Impact of training set size on object-based land cover classification: A comparison of three classifiers. International Journal of Applied Geospatial Research 5(3): 49–67. DOI: 10.4018/ijagr.2014070104.

, et al. (2019) Mapping with confidence; delineating seagrass habitats using unoccupied aerial systems (UAS). Remote Sensing in Ecology and Conservation 5(2): 121–135. DOI: 10.1002/rse2.98.

53.

Öberg

(2006) Primary production by macroalgae in Kattegat, estimated from monitoring data, seafloor properties, and model simulations. Chemical Society Reviews 26(19): 2415–2432. DOI: 10.1016/j.csr.2006.07.005.

54.

Olesen

Sand-Jensen

(1994) Patch dynamics of eelgrass Zostera marina. Marine Ecology Progress Series 106(1–2): 147–156. DOI: 10.3354/meps106147.

55.

Olesen

(1999) Reproduction in Danish eelgrass (Zostera marina L.) stands: size-dependence and biomass partitioning. Aquatic Botany 65(1–4): 209–219. DOI: 10.1016/S0304-3770(99)00041-8.

, et al. (2006) A global crisis for seagrass ecosystems. BioScience 56: 987. DOI: 10.1641/0006-3568(2006)56[987:AGCFSE]2.0.CO;2.

, et al. (2012) Seed addition facilitates eelgrass recovery in a coastal bay system. Marine Ecology Progress Series 448: 177–195. DOI: 10.3354/meps09522.

, et al. (2020) Restoration of seagrass habitat leads to rapid recovery of coastal ecosystem services. Science Advances 6(41): 1–9. DOI: 10.1126/sciadv.abc6434.

, et al. (2018) Kortlægning Af Ålegræsenge Med Ortofotos (Danish). Roskilde: Scientific ReportDCE, 265, pp. 1–70.

, et al. (2009) Odense pilot river basin: implementation of the EU water framework directive in a shallow eutrophic estuary (Odense Fjord, Denmark) and its upstream catchment. Hydrobiologia 629(1): 71–89. DOI: 10.1007/s10750-009-9774-0.

(2012) Effects of filamentous macroalgae mats on growth and survival of eelgrass, Zostera marina, seedlings. Aquatic Botany 99: 41–48. DOI: 10.1016/j.aquabot.2012.01.005.

, et al. (1999) Modelling advective transport of Ulva lactuca (L) in the sheltered bay, Møllekrogen, Roskilde Fjord, Denmark. Hydrobiologia. 397: 241–252. DOI: 10.1023/A:1003790625535.

, et al. (2011) Extinction risk assessment of the world's seagrass species. Biological Conservation 144: 1961–1971. DOI: 10.1016/j.biocon.2011.04.010.

, et al. (2019) Blue carbon stock in Zostera noltei meadows at Ria de Aveiro coastal lagoon (Portugal) over a decade. Scientific Reports 9(1): 1–13. DOI: 10.1038/s41598-019-50425-4.

, et al. (unpub). Working title: Identifying stress-impact on a coastal eelgrass meadow over a 6-year period – Application of UAV-based imagery and Object-Based Image Analysis.

, et al. (2016) Droner–et nyt værdifuldt værktøj til kystnær overvågning (danish). Vand and Jord, E-Bog Nr 3: 125–129.

(2018) Marin monitering ved brug af droner: Ålegræs-og presfaktor-kortlægning ved billedanalyse (danish). In: Conference Contribution. Danish Marine Science Meeting, Odense, Denmark.

68.

SSI (2019) SSI. USA: SSI. Available at: http://www.sigmaplot.co.uk/products/sigmaplot/sigmaplot-details.php (accessed 04 April 2019).

69.

Stæhr

Groom

Krause-Jensen

, et al. (2019) Use of Remote Sensing Technologies for Monitoring Chlorophyll a and Submerged Aquatic Vegetation in Danish Coastal Waters Technical Report No. 139. Roskilde: DCE, pp. 1–66.

70.

The ministry of traffic and buildings , 2016. Act of Drone-Flight in Urban Areas (Danish). The Ministry of Traffic and Buildings: 1–14.

71.

Tanner

(2005) Edge effects on fauna in fragmented seagrass meadows. Austral Ecology 30(2): 210–218. DOI: 10.1111/j.1442-9993.2005.01438.x.

72.

The Agency of Data Distribution and Efficiency , 2021. DTK/Skaermkort. Downloaded May 2021.

73.

The ministry of traffic and buildings , 2017. Act of Drone-Flight Outside Urban Areas (Danish). The Ministry of Traffic and Buildings: 1–7.

74.

Trimble Inc (2020). eCognition Developer 9.3 – 9.5. Available at: http://www.ecognition.com (accessed 01 October 2020).

75.

Tzotsos

Argialas

(2008) Support vector machine classification for object-based image analysis. In: Blaschke

Lang

Hay

(eds.), Object-Based Image Analysis – Spatial Concepts for Knowledge-Driven Remote Sensing Applications. Berlin Heidelberg, Germany: Springer-Verlag, pp. 663–678.

76.

Valdemarsen

Canal-Verges

Kristensen

, et al. (2010) Vulnerability of Zostera marina seedlings to physical stress. Marine Ecology Progress Series 418: 119–130. DOI: 10.3354/meps08828.

77.

Valdemarsen

Wendelboe

Egelund

, et al. (2011) Burial of seeds and seedlings by the lugworm Arenicola marina hampers eelgrass (Zostera marina) recovery. Journal of Experimental Marine Biology and Ecology 410: 45–52. DOI: 10.1016/j.jembe.2011.10.006.

78.

Valdemarsen

Quintana

Kristensen

, et al. (2014) Recovery of organic-enriched sediments through microbial degradation: Implications for eutrophic estuaries. Marine Ecology Progress Series 503: 41–58. DOI: 10.3354/meps10747.

79.

Ward

Kemp

Boynton

(1984) The influence of waves and seagrass communities on suspended particulates in an estuarine embayment. Marine Geology 59(1–4): 85–103. DOI: 10.1016/0025-3227(84)90089-6.

80.

Warry

Hindell

Macreadie

, et al. (2009) Integrating edge effects into studies of habitat fragmentation: a test using meiofauna in seagrass. Oecologia 159(4): 883–892. DOI: 10.1007/s00442-008-1258-9.

81.

Waycott

Duarte

Carruthers

TJB

, et al. (2009) Accelerating loss of seagrasses across the globe threatens coastal ecosystems. Proceeding of the National Academy of Sciences of the united states of America 106(30): 12377–12381. DOI: 10.1073/pnas.0905620106.

82.

Wicks

Koch

O’Neil

, et al. (2009) Effects of sediment organic content and hydrodynamic conditions on the growth and distribution of Zostera marina. Marine Ecology Progress Series 378: 71–80. DOI: 10.3354/meps07885.

83.

Williams

(2007) Introduced species in seagrass ecosystems: Status and concerns. Journal of Experimental Marine Biology and Ecology 350(1): 89–110. DOI: 10.1016/j.jembe.2007.05.032.