High-Resolution Three-Dimensional Mapping of Eelgrass (Zostera marina) Habitat and Blue Carbon Using Drone-Borne LiDAR

Highlights

What are the main findings?

• Digital models from drone-borne LiDAR data provide bathymetry and 3D structure of an eelgrass bed.
• These models allow quantification of eelgrass bed habitat volume and living tissue carbon stock in 3D.

What are the implications of the main findings?

• Drone-borne LiDAR presents a novel tool for assessing submerged vegetated habitat structure and carbon storage.
• Drone-borne LiDAR improves accessibility and resolution versus airborne LiDAR, and can supplement other remote-sensing approaches with high-resolution, 3D data.

Abstract

The accessibility of flying drones (unmanned aerial vehicles) presents reproducible and cost-effective methods to monitor submerged aquatic vegetation. In particular, drone-borne topobathymetric LiDAR provides high-resolution (cm-scale), three-dimensional information about the geometry and structure of surveyed areas, allowing for quantification of vegetation volume in addition to bathymetry. For seagrasses, this information can advance research regarding the structure of canopies in relation to blue carbon storage and biodiversity. Here, we demonstrate how drone-borne LiDAR can be used to estimate the habitat volume of eelgrass (Zostera marina) within a sheltered bay in Norway. After classifying LiDAR points using a Random Forest model, we created a Digital Terrain Model of the sea floor and a Digital Surface Model of the eelgrass canopy. From these models, we showed that eelgrass canopy volume can be estimated (between 862 and 1099 m³ across the small study area) and the above-ground carbon stock in living tissue can be quantified (between 96 and 122 kg C). To our knowledge, this is the first study to utilise drone-borne LiDAR to quantify the habitat volume and carbon-storage potential of a marine habitat-forming species like eelgrass, demonstrating a novel methodology for providing reproducible and high-resolution data of submerged aquatic habitats.

1. Introduction

The ecosystem services provided by seagrasses are numerous and increasingly important in the context of rapid climatic change and widespread habitat degradation [1]. Their high primary productivity and ability to trap and retain organic matter show that seagrass canopies and sediments are globally important for storage of blue carbon (i.e., organic carbon stored in biomass, soil, and sediments of marine coastal vegetated ecosystems) [2,3,4,5], hence their formal inclusion in 2013 into the IPCC’s Guidelines for Greenhouse Gas Inventories for Wetlands [6]. The presence of seagrass beds has also demonstrated enhanced protection from coastal erosion [7]. As habitat-forming species, seagrasses provide structural complexity that supports diverse benthic infauna and epifauna, offering nursery and refuge areas for invertebrates and fishes, as well as sources of food for megaherbivores and predators [8,9,10,11]. These habitat provisions have been shown to support the basis of fisheries productivity globally [12,13,14].

Seagrasses are experiencing widespread global decline and degradation [15,16]. In recognition of this, seagrass protection and restoration was included in the Conservation of Migratory Species (CMS) resolution at the 14th United Nations Conference of the Parties (COP14) in 2024 [17]. This resolution focuses on the need for monitoring of seagrass habitat and associated biodiversity, particularly in relation to known drivers of seagrass loss [17]. The fine-scale monitoring of seagrass has traditionally relied on in situ observation and manual data collection, such as snorkel or SCUBA-based surveys. In general, these approaches are often labour-intensive, costly, and restricted to easily accessible areas [18]. Large-scale seagrass monitoring often relies on satellite-derived data products, which provide the widest geographical coverage, but at the cost of lower-resolution data (e.g., several m pixel resolution for publicly available data, ≤1 m for commercial data), with limited differentiation between species and habitat types [19]. Aircraft-based seagrass monitoring methods have also been developed, increasing data resolution (<1 m pixel resolution), but simultaneously increasing costs and decreasing temporal coverage due to logistical constraints of deploying survey aircraft (relative to consistent satellite coverage) [20].

Alternatively, flying drones (technically referred to as unmanned aerial vehicles, UAVs) have proven effective to expedite a non-invasive, reproducible and cost-efficient method for habitat monitoring across a variety of habitats and conditions [21,22]. For example, drones mounted with optical sensors have successfully been used to map submerged vegetated habitats including seagrass and macroalgae by differentiating species and habitats based on their unique optical signatures [23,24]. As they are relatively easy to deploy, flying drones can facilitate on-demand and/or repeated surveys, often at the highest remotely sensed resolutions (centimetre-scale) [21].

Air-borne Light Detection and Ranging (LiDAR) systems carried on small, crewed aircraft have been applied to map terrestrial and marine ecosystems in three-dimensional space [25,26,27], particularly used in forest research [28]. More recently, drone-mounted systems now offer similar capabilities at lower cost and higher spatial precision, although with a somewhat more limited spatial range [27]. Typically, drone-borne LiDAR-collected point cloud densities are at a scale of some hundred points per square metre, while crewed aircraft collect a point density one magnitude lower.

LiDAR systems work by emitting pulsed laser beams, whereby the time it takes for the laser to reflect off a surface and return to the sensor is used to calculate the distance, and thus the elevation of the targeted surface [29]. The resulting point cloud data therefore captures the geometry and height of surfaces within the scanned area [30]. Topobathymetric LiDAR surveys land, water, and submerged surfaces by using a green laser (i.e., wavelength = 532 nanometres) that penetrates the water column down to 1–3 Secchi depths, depending on water clarity and the power of the LiDAR system [31]. The underwater elevation of points in topobathymetric LiDAR therefore corresponds to the elevation of the reflected surface relative to the water surface. This is the case as these systems are optimized to detect changes in the green laser beam signal at the air–water interface. There, part of the laser beam is reflected back to the LiDAR receiver, while other parts of the laser beam continue to propagate through the water column at slower speeds (Figure 1) [32]. In summary, topobathymetric LiDAR generates accurately georeferenced point clouds that provide detailed elevation profiles of both the sea floor and sea floor cover, i.e., submerged vegetation, facilitating geometric investigation of their structure.

Not only does LiDAR capture geometric information (e.g., x, y, z), it also captures radiometric data [33], including the intensity, or return strength, of the reflected laser beam [34]. The scale of intensity represents the proportion of light reflected from a surface, which varies depending on the material and texture of the target. This variation has been used to differentiate between ground structures and vegetation [35], and also to distinguish between different species of terrestrial plants [36]. Overall, the inclusion of intensity values alongside geometric data greatly enhances LiDAR application, especially to distinguish vegetation from natural substrates or artificial surfaces [37,38].

Recent advances in drone-borne LiDAR technology have made medium-sized drones increasingly accessible to a wide variety of users, including researchers and managers. However, drone-borne topobathymetric LiDAR systems (i.e., hardware plus software) remain few. Currently, some available systems offer dedicated software for automated post-processing, including georeferencing, water surface detection, and point cloud intensity correction at ‘survey-grade’ quality (e.g., precision and accuracy ≤ 3 cm from up to 100 m altitude, according to the manufacturer) [39]. Such accessibility of drone technology, paired with user-friendly data handling, can enable cost-effective and reproducible methods necessary to meet global ambitions of seagrass monitoring. Moreover, the geometry of LiDAR data enables three-dimensional quantification and visualisation of surveyed areas at high resolution, often at scales of a few centimetres. These high-resolution data can serve as ground-truth data for lower-resolution remote-sensing data, facilitating upscaling across the ‘observational pyramid’ [24,40]. At the fine scale, LiDAR data can enhance our understanding of how the structure of seagrass canopies and meadows relates to their ecological functions, including habitat provisioning, hydrodynamics, and physical coastal conditions, as well as biomass and blue carbon stock content and capacity. To the authors’ knowledge, no published studies have yet demonstrated the three-dimensional quantification of seagrass habitat using drone-borne LiDAR.

In this study, we present a novel remote-sensing approach to identify and map shallow-water seagrass habitats, and to quantify the standing biomass and canopy carbon stock (Figure 2). Using a medium-sized drone equipped with a compact green LiDAR sensor (12.3 kg including sensor), we demonstrate how geometric and radiometric point cloud data can provide new insight on benthic habitats. Using a Random Forest (RF) model, we classified LiDAR point cloud data and created a Digital Terrain Model (DTM) of the sea floor bathymetry and a Digital Surface Model (DSM) of the vegetation canopy cover. Together, these models provide high-resolution volumetric information and three-dimensional visualisation of a seagrass habitat area at a study site in Norway, enabling remotely sensed quantification of above-ground biomass and associated canopy carbon stock.

2. Materials and Methods

2.1. Study Species and Area

Eelgrass (Zostera marina) is a common seagrass species found throughout the northern hemisphere [41]. In Europe, its distribution spans from sub-tropical conditions in southern Iberia to Arctic conditions in northern Norway [42]. Our study area was located in a sheltered, shallow bay containing eelgrass meadows at Ølbergholmen, a small peninsula in Vestfold County, south-eastern Norway (59°N, 10.13°E, Figure 3).

2.2. LiDAR Data Collection

Data were collected using a topobathymetric LiDAR sensor (Navigator, YellowScan, Saint-Clément-de-Rivière, France) mounted on a midrange UAV (Tundra 2 equipped with Endurance rotor arms, Hexadrone, Saint-Ferréol-d’Auroure, France) [39]. This system includes a laser with an emitting wavelength of 532 nm, which achieved a horizontal accuracy of 2.4 cm root mean square error (RMSE) and a precision of 0.5 cm (mean standard deviation, S.D.) via calibration (above water) performed by the manufacturer approximately 1 month before deployment. The system has a constant beam angle of 20° to either side of nadir, resulting in a 40° field of view for scanning [39]. The first of two drone flights was conducted on 29 August 2024 at 05:35 h, at an elevation of 50 m, with 50% overlap between flight lines as recommended by the manufacturer. A second flight occurred later the same day at 06:18 h, at 25 m elevation, also with 50% overlap. Note that both LiDAR flights occurred during a period when eelgrass was fully submerged, i.e., no leaves floating on the water surface. In subsequent analyses, we compare LiDAR-derived eelgrass canopy metrics at both altitudes to assess the effect of drone elevation, and the resulting point cloud density, on canopy volume estimates.

2.3. In Situ Data Collection

To validate LiDAR-generated canopy height and ground elevation values, eelgrass canopy height and ground elevation both above and below water were collected in situ on 2 November 2024 (Figure S1). As in situ measurements were taken after the LiDAR flights, to assess potential seasonal changes in canopy height, we compared canopy measurements from a previous campaign to the same area in September 2023 to data from November 2024. We expected a reduced height in November, which is beyond the period of maximum eelgrass height, which typically occurs around September in the region [43]. Canopy height was measured visually using mask and snorkel and a measuring stick, with height corresponding to the observer’s estimated average eelgrass canopy height (in metres), of undisturbed leaves, within a 20 cm radius of the measuring stick. Ground elevation data were gathered using an Emlid (Budapest, Hungary) Reach RS3 real-time kinematic (RTK) positioning Global Navigation Satellite System (GNSS) receiver [44]. Elevation values under water were taken on sandy substrates only (Figure S1).

2.4. LiDAR Bathymetry Comparison with Sonar

For deeper areas (>1.5 m) that could not be accessed with the pole-mounted handheld GNSS, bathymetry data were collected using a BioSonics MX 200 kHz single-beam aquatic habitat mapping echosounder, installed on an uncrewed surface vehicle (USV, Maritime Robotics Otter Pro). This small USV (2 m long, 70 kg) was deployed across the study area using vehicle guidance to follow a pre-programmed mission ‘lawnmower’ pattern, at approximately 1 m/s with 5–10 m cross-track spacing (Figure S1), utilising an SGB Ellipse-D Internal Navigation System (INS) and dual GNSS antennas. The echosounder was mounted on the USV’s bow, with a 9° conical beam angle transducer pointed at the sea floor and connected to BioSonics MX acquisition electronics contained within the vehicle’s payload. The system logged the transducer signals along with time, latitude, and longitude from the vehicle’s INS. Echosounder data were collected on 13 September 2023 at 08:48 h, and data processing was performed in BioSonics’ Visual Aquatic software version 1.0.0.13146 [45]. Considering variation in tide between sampling periods (tidal amplitude in the area is <0.3 m), we adjusted the echosounder data to match the tide level occurring during LiDAR sampling.

2.5. Raw LiDAR Data Post-Processing

LiDAR data were refined following the flight trajectory and georeferenced using YellowScan’s CloudStation software version 2507.0.1-b6dd9e0c5 [46], with point cloud processing based on the equations in [47]. Refinement, georeferencing, and accuracy assessment were performed by accessing GNSS base stations to account for drone position and altitude, plus an Inertial Measurement Unit (IMU) contained in the LiDAR sensor that accounts for scanner range and angle [48]. The resulting point cloud was georeferenced to the local coordinate system (UTM Zone 32N, ETRS 1989) and vertical reference frame (NN2000) [49]. Vertical and horizontal position accuracy estimates were calculated via GNSS Post-Processed Kinematic (PPK) [46]. The full reports of point cloud processing in CloudStation are available for viewing in the electronic Supplementary Materials.

The initial point classification was conducted using the automated Terrain Classification plug-in module in CloudStation [46]. This included classification of points originating from one of four classes: land (i.e., all points outside of the water), the water surface, water column noise, and sea floor (Figure S2A). The point cloud was also colourised in CloudStation using the software’s Colorization Mode, from RGB data collected simultaneously by the Navigator’s onboard 1-megapixel RGB camera (n = 218 images) [46]. Additional point attributes included scan angle, scan return number, and intensity. Raw intensity values were corrected during post-processing in CloudStation using the bathymetric LiDAR equation [47,50], scaling intensity data to a 16-bit range, reported on a scale between 32 and 8518 [39,46].

2.6. Data Cleaning

While automatic point cloud classification in CloudStation expedited raw post-processing (Figure S2A), further refinement was required to distinguish between true sea floor (including both sandy and rocky bottoms) and sea floor cover (i.e., vegetation) (Figure S2B). Following raw post-processing, the point cloud was loaded into the open-source software CloudCompare version 2.13.2 [51] for inspection and further manual processing. We investigated classified sea floor and water column noise points and found that some points classified as water column noise in CloudStation originated from benthic vegetation, which was confirmed via visual inspection (see Figure S2B). This was likely due to the elevation at which vegetation was suspended within the water column relative to the sea floor. Filtering by intensity values (see next section), we retained points originating from vegetation and manually removed the remaining water column noise points, which were easily identified by visual inspection from the point cloud cross-sections (Figure S2B). Final cleaning was also performed in CloudCompare, where areas lacking sufficient sea floor point density (due to high water column noise) were removed to ensure reliable creation of digital models in subsequent analyses.

2.7. Point Cloud Classification

In the present and subsequent sections, the code written for all analyses in R (version 4.5.0) [52] was assisted by the Claude Sonnet 3.5 Large Language Model [53]. To be able to create a dataset for classifying all points in a Random Forest model, we first manually annotated the point cloud by identifying vegetation and sea floor points within the point cloud, based on ground-truth observations. Manual inspection of the point cloud’s intensity values, at the location of ground-truth data points (Figure S1), showed a marked difference between ‘true’ sea floor and vegetation intensity values (Figure S2B). Based on the location of eelgrass ground-truth points, we extracted the range, mean, and standard deviation of eelgrass points’ intensity values within the point cloud. We then annotated the point cloud data by labelling all points with intensity values of less than the mean intensity value of eelgrass plus one standard deviation as vegetation, and points higher than this range as sea floor. This annotated dataset was used to train a Random Forest (RF) model for point classification.

Based hereon, we created a probability-based prior, with points within the pre-defined vegetation intensity range defined as >90% probability of being vegetation and points outside of this range as <10% probability. In addition to the probability-based prior, we used red, green, and blue values (each ranging from 0 to 255, equivalent to 8-bit) derived from the simultaneously captured RGB imagery, as well as the nearest-neighbour distance (in centimetres) of points as predictor variables in the RF model to predict points as either sea floor or vegetation. The model was created using the randomForest package in R [54]. The number of trees (ntree) was determined by the minimum Out of Bag (OOB) error across a range of 10 to 100 trees. The number of randomly drawn candidate variables was determined by grid search using the caret package [55]. Data were split into 80% training and 20% testing subsets to evaluate final model performance via confusion matrix and precision scores. OOB error and variable importance (as measured by Mean Decrease Accuracy, or the reduction in model accuracy when a variable is randomly permuted) were extracted using the randomForest package [54]. After assessment, the resulting RF model was used to classify all points using the base R predict function [52].

2.8. Digital Terrain and Surface Models

Following point cloud classification, we created a DTM from the sea floor points and a DSM from the vegetation points using the lidR R package [56,57]. The sea floor DTM was created using the Inverse Distance Weighting (IDW) algorithm, using 10 nearest-neighbour points and a weighting distance parameter of 2 (i.e., squared distance). The DTM was then smoothed in order to reduce the effects of fine-scale features of the bathymetry model that likely resulted from point cloud noise at the vegetation/sea floor interface [58]. Smoothing was completed using the focal function from the terra package in R [59], applying the mean across a moving window to ensure smoothness. We performed a sensitivity analysis by testing different smoothing window (i.e., neighbourhood) values from low to high to investigate the effects on canopy height results.

The vegetation DSM was created using the point-to-raster algorithm in the lidR package, which attributes the elevation of the highest point within each raster cell [56,57]. As an objective of this study was to quantify eelgrass habitat volume, to prevent excessive voids within the modelled vegetation canopy DSM, we replaced each point with a ‘disk’ of 8 points surrounding the original point. This approach has been demonstrated to improve terrestrial forest canopy height models (CHMs) by removing ‘data pits’, and as such, the routine creates a DSM that captures the ‘outside’ of the canopy, thereby creating a smoothing effect on the DSM that aligns with the human perception of a canopy cover [60]. We assumed this would best align with the seagrass canopy height measurements. To assess the sensitivity of this smoothing approach, we tested disk radii (i.e., smoothing window sizes) from 0.1 to 0.5 m at 0.1 m increments and evaluated their effects on estimated canopy height and volume.

Following the creation of digital models, we restricted the vegetation DSM to areas of eelgrass manually using ground-truth observations, in combination with visual assessment of the colourized LiDAR point cloud data and expert knowledge of the study area. Also, final DTM and DSM rasters were masked to overlapping extents between the 50 m and 25 m flight datasets, allowing for comparison. After restricting the DSM to areas of eelgrass, we calculated the cell-specific canopy height (CH_r, in metres) as the difference between the DSM and DTM elevations (z) of cell r:

$${\mathit{C}\mathit{H}}_{\mathit{r}}={\mathit{z}}_{{\mathit{D}\mathit{S}\mathit{M}}_{\mathit{r}}}-{\mathit{z}}_{{\mathit{D}\mathit{T}\mathit{M}}_{\mathit{r}}}$$

(1)

We then calculated the volume of eelgrass habitat across the study area (in cubic metres) as the sum of the volume of all canopy cells (n), calculating the volume of cell r by multiplying canopy height (CH_r, in metres) by the area of each cell (in square metres):

$$\mathit{V}\mathit{o}\mathit{l}\mathit{u}\mathit{m}\mathit{e}={\sum }_{\mathit{r}=1}^{\mathit{n}}{\mathit{C}\mathit{H}}_{\mathit{r}}\times {\mathit{a}\mathit{r}\mathit{e}\mathit{a}}_{{\mathit{C}\mathit{H}}_{\mathit{r}}}$$

(2)

2.9. Validation

We validated the LiDAR-derived elevation values for both terrestrial (above-water) and submerged (below-water) surfaces by comparing them with field-measured ground-truth data. Validation metrics included mean error, S.D., and RMSE between GNSS ground-truth data (handheld and echosounder) and LiDAR values. In situ eelgrass canopy height was compared to LiDAR vegetation canopy height (i.e., the elevation difference between vegetation-classified LiDAR points and the DTM below the point). For all validations, we found the nearest LiDAR point to each ground-truth coordinate, based on Euclidean distance, using the nn2 function from the RANN package in R [61].

2.10. Eelgrass Biomass and Carbon Content

We calculated eelgrass biomass and organic carbon content from the LiDAR-derived canopy model (i.e., Equations (1) and (2)) by incorporating the mean biomass per unit area calculated at the study site. This value was derived from eelgrass samples (n = 42) collected in September 2023, by manually measuring the height and wet weight of eelgrass in 20 × 20 cm cells (n = 42, Table S1) [62]. The biomass per unit area (hereafter referred to as volumetric biomass, B_vol, g WW m⁻³) was then calculated using the equation:

$${\mathit{B}}_{\mathit{v}\mathit{o}\mathit{l}}={\displaystyle \frac{\mathit{W}\mathit{W}/\mathit{A}}{\mathit{H}}}$$

(3)

where WW equals the wet weight (grams) of eelgrass samples, A equals the sample cell area (square metres), and H equals the mean measured height of eelgrass (metres) in the sample cell. We also report biomass in areal units, i.e., g WW m⁻² (WW/A). From the wet weight biomass values per unit area, we converted to dry weight (DW) using a conversion value derived from subsamples of the in situ samples mentioned above (n = 85, DW/WW = 0.24, Table S2), and from dry weight to carbon (C/DW = 0.34) using a standard C/DW conversion value from literature [63,64]. In subsequent analysis, we assumed that this calculated mean was representative of the eelgrass canopy at the time of LiDAR data sampling (August 2024). We calculated total eelgrass biomass (WW, kilograms) across the study area using Equation (4):

$$\mathit{B}\mathit{i}\mathit{o}\mathit{m}\mathit{a}\mathit{s}\mathit{s}=\left(\mathit{V}\mathit{o}\mathit{l}\mathit{u}\mathit{m}\mathit{e}\times {\mathit{B}}_{\mathit{v}\mathit{o}\mathit{l}}\right)\times 0.001$$

(4)

where Volume equals the LiDAR-derived eelgrass habitat volume (in cubic metres), for both the 25 m and 50 m datasets, and B_vol equals the value calculated from Equation (3). Using the conversion ratios, we converted total biomass WW (kilograms) to biomass DW (kilograms), and C (kilograms). For reporting purposes, we also summarised all biomass values per square metre. We recognise that a potential error could be introduced to the final estimated biomass and C content due to the temporal mismatch between the LiDAR and the in situ sampling. However, we assume this error to be relatively low (±10–20%), as DW/WW and C/DW ratios in fresh eelgrass are relatively constant [65,66]; moreover, the main aim of this paper is intended to demonstrate the potential of the methodology rather than providing a precise biomass measurement.

3. Results

3.1. LiDAR-Derived Bathymetry DTM and DSM

For the 50 m LiDAR data, a total of 4,356,280 points were recorded across the 32,676 m² study area following raw data post-processing (Figure 4A), with an average horizontal and vertical accuracy of 1 ± 0.4 cm (RMSE = 1.1 cm) according to PPK (see electronic Supplementary Materials). After point cloud cleaning, we retained 895,979 points across an area of 12,868 m², with a mean density of 70 points per 1 m², 61 points per 1 m³ voxel (volume pixel), and a mean nearest-neighbour distance of 8.9 cm (±<0.01 S.E.) (Figure 4B). From ground-truth data, points between 32 (minimum value) and 1263 intensity were annotated as vegetation points (eelgrass points mean intensity = 863 + 400 S.D., Figure 4C), with points above this range were annotated as sea floor. The 25 m LiDAR dataset initially contained 5,682,638 points across 23,904 m² (Figure 4D), with an average horizontal accuracy of 1.6 cm ± 0.2 cm (RMSE = 1.6 cm) and an average vertical accuracy of 1.7 cm ± 0.2 cm (RMSE = 1.7 cm, see electronic Supplementary Materials). After cleaning, 2,112,557 points remained across 16,840 m² with a mean density of 125 points per 1 m², 105 points per 1 m³ voxel, and a mean nearest-neighbour distance of 6.9 cm (±<0.01 S.E.) (Figure 4E). Points annotated as vegetation were those with intensity values from 100 (minimum value) to 1810 (eelgrass points mean intensity = 1512 + 298 S.D., Figure 4F), and points above this range were annotated as sea floor. The depth range of both cleaned datasets extended from 0 cm (i.e., the land–water interface) down to approximately 2.5 m.

From the resulting point clouds, the RF models used for point cloud classification showed 100% accuracy in distinguishing between sea floor and vegetation within our test datasets. Both the 50 m and 25 m models showed low OOB error and classification error rates, converging to 0 before reaching 40 trees (Figure S3A,B). The intensity-based probability-prior was by far the most important variable in predicting point classification (Figure S3C,D). The final RF models were then used to classify all points as either sea floor or vegetation, allowing generation of the DTMs and DSMs.

The DTMs and DSMs from both 50 m and 25 m LiDAR datasets were created at a resolution of 10 cm, which is slightly coarser than the average point spacing of 8.9 cm in the 50 m dataset. We chose a large smoothing window (neighbourhood = 87) for DTM smoothing, as the larger window resulted in less noise present in subsequent canopy height results (Figure S4). The vegetation DSMs were restricted to areas of eelgrass, removing canopy coverage of macroalgae species (primarily Fucus serratus and Fucus vesiculosus) growing on hard substrate surrounding a bedrock outcropping in the southwestern corner of the bay (Figure 4 and Figure 5). Subtracting the DTM from the DSM resulted in the eelgrass canopy height model (Figure 5).

Comparing LiDAR-derived sea floor elevations with in situ RTK GNSS measurements (RTK GNSS mean RMSE of Eastings, Northings, and elevation of 1.5 (±1.1), 1.8 (±3.3), and 2.4 (±3.4) cm, respectively) demonstrated a strong consistency between the 50 m LiDAR dataset and ground-truth elevation, showing a mean vertical error of only 0.5 cm (12.0 cm S.D., 11.7 cm RMSE) (

Table 1. LiDAR-derived terrestrial elevation, bathymetry elevation, and eelgrass canopy height, relative to ground-truth values. Eelgrass canopy height was compared to manual measurements, and bathymetry was compared to handheld Global Navigation Satellite System (GNSS) elevation measurements and echosounder-derived bathymetry, while terrestrial elevation (outside of the focal study area) was compared with handheld GNSS measurements only. Validation is reported as the mean vertical error, standard deviation (S.D.) and root mean square error (RMSE) in centimetres between LiDAR elevation values and ground-truth values.

	Validation Relationship	Mean Error (cm)	S.D. (cm)	RMSE (cm)
Terrestrial elevation	25 m LiDAR ~ Handheld GNSS	0.6	3.5	3.5
	50 m LiDAR ~ Handheld GNSS	0.6	6.8	6.7
Bathymetry elevation	25 m LiDAR ~ Handheld GNSS	6.5	15.4	16.3
	25 m LiDAR ~ Echosounder	10.0	8.9	13.4
	50 m LiDAR ~ Handheld GNSS	0.5	12.0	11.7
	50 m LiDAR ~ Echosounder	13.5	7.4	15.4
Eelgrass canopy height	25 m LiDAR ~ manual measurement	9.8	8.1	12.4
	50 m LiDAR ~ manual measurement	13.2	5.4	14.2

, Figure 6). The 25 m dataset showed a larger offset than the 50 m dataset (mean vertical error of 6.5 cm, 15.4 cm S.D., 16.3 cm RMSE) (Table 1, Figure 6). When compared with echosounder-derived sea floor bathymetry, the 25 m LiDAR dataset showed better agreement than the 50 m dataset, with a 10 cm mean vertical error (8.9 cm S.D., 13.4 cm RMSE) (Table 1, Figure 6). For terrestrial LiDAR points (i.e., on land outside of the study area), both datasets produced similar vertical accuracy (mean error = 0.6 cm), with the 25 m dataset displaying slightly lower S.D. (3.5 cm) and RMSE (3.5 cm) (Table 1, Figure S5).

3.2. Seagrass Canopy Height and Meadow Volume

Canopy heights derived from both the 50 m and 25 m LiDAR datasets (from late August 2024) were generally lower than manually measured ground-truth data (from early November 2024) (Figure 7A). Comparing the eelgrass canopy heights obtained by manual measurements between November 2024 and September 2023, canopy heights were slightly higher in September 2023 than in November 2024, with a strong positive relationship between the two measurements (R² = 0.37), barring one outlier (Figure S6).

Modelled eelgrass canopy height, area, and volume were influenced by both flight elevation and the smoothing radius of the DSM (Figure 7A,B). Total eelgrass canopy area and volume were higher for the 25 m dataset, while increasing the DSM smoothing window increased canopy area and volume for both datasets (Figure 7B). Canopy area ranged from 4723 to 5703 m² (mean = 5287, S.D. = 381) and volume from 818 to 1099 m³ (mean = 978, S.D. = 109) across smoothing radii for the 25 m dataset, while area ranged from 3881 to 4909 m² (mean = 4484, S.D. = 398) and volume from 629 to 862 m³ (mean = 763, S.D. = 90) for the 50 m dataset (Figure 7B). With the largest smoothing radius (0.5 m), canopy height ranged from 0.0 to 0.98 m for the 25 m dataset (Figure 7C), and from 0.0 to 0.64 m for the 50 m dataset (Figure 7D). The distribution of canopy height cells displayed a general increase concurrent with an increase in smoothing radius (Figure S7). When comparing canopy height and bathymetry elevation between the 50 m and 25 m datasets (Figure S8A), that of the 50 m dataset was generally lower, according to the negative-skew of the comparative histograms (Figure S8B). Moreover, the areal coverage of eelgrass canopy height raster cells was higher within the 25 m dataset (Figure S9).

3.3. Above-Ground Eelgrass Biomass and Carbon

The mean in situ eelgrass volumetric biomass equalled 1363.5 g WW m⁻³ (S.D. = 999.1) (

Table 2. Results deriving the mean (and standard deviation, S.D.) biomass per unit area from samples collected at the study site (n = 42) in September 2023 (see Table S1). This includes the biomass density in grams (g) of wet weight (WW) m⁻² and m⁻³, the biomass density in g dry weight (DW) m⁻² and m⁻³, and biomass in g carbon (C) m⁻² and m⁻³. Wet weight to dry weight ratios were derived from subsamples (DW/WW = 0.24, Table S2), while a standard dry weight to carbon ratio was used from the literature (C/DW = 0.34).

	Biomass (g WW m−2)	Biomass Density (g WW m−3)	Biomass (g DW m−2)	Biomass Density (g DW m−3)	Biomass (g C m−2)	C Density (g C m−3)
Mean (S.D.)	364.1 (343.7)	1363.5 (999.1)	86.3 (81.5)	323.1 (236.8)	27.6 (26.1)	103.4 (75.8)

and Table S1). Applying this value, we then calculated the total eelgrass biomass (total kg WW, kg WW m⁻²) across the study area for both the 25 m and 50 m datasets (DSM smoothing radius = 0.5 m,

Table 3. Results summarising eelgrass biomass across the study area, by applying the mean volumetric biomass (B_vol) value, to both the 25 m and 50 m LiDAR-derived eelgrass canopy models (DSM smoothing radius of 0.5 m only). Included is the total biomass in kilograms (kg) in wet weight (WW) across the study area, the biomass (kg) WW m⁻², biomass dry weight (DW) across the study area, biomass DW m⁻², biomass (kg) of carbon (C) across the study area, and biomass (kg) C m⁻².

LiDAR Dataset		Mean Bvol
25 m	Biomass kg WW (study area)	1499.0
	Biomass kg WW m−2	0.263
	Biomass kg DW (study area)	359.8
	Biomass kg DW m−2	0.063
	C kg (study area)	122.3
	C kg m−2	0.021
50 m	Biomass kg WW (study area)	1174.7
	Biomass kg WW m−2	0.239
	Biomass kg DW (study area)	281.9
	Biomass kg DW m−2	0.057
	C kg (study area)	95.9
	C kg m−2	0.019

). After converting to carbon content (via DW, see Methods), we estimated 122.3 kg C from the 25 m dataset and 95.9 kg C from the 50 m dataset (Table 3). This equates to approximately 19.5 g C m⁻² (50 m dataset) or 21.4 g C m⁻² (25 m dataset) (Table 3).

4. Discussion

From drone-borne topobathymetric LiDAR data, we were able to estimate the areal extent and the specific canopy height of a seagrass meadow habitat within a confined coastal study area at a spatial resolution of 10 cm. By creating high-resolution digital models of both the sea floor and the overlaying seagrass canopy, the habitat volume, biomass, and above-ground carbon stock were estimated. These novel results build upon previous applications of airborne topobathymetric LiDAR for mapping the spatial extent of marine vegetation and classifying habitat types [67,68,69,70,71]. However, this case study advances these applications by additionally deriving and analysing the three-dimensional geometric structure of submerged vegetation, which enables quantification of habitat volume and seagrass biomass and provides remotely sensed estimates of blue carbon stocks.

Airborne LiDAR has been applied extensively in terrestrial forest research [72]. For example, LiDAR-derived forest canopy structure has been related to shading and microclimatic conditions—factors that directly influence plant growth, competition, and community composition [73,74]. Airborne LiDAR also provides accurate tree height and crown area measurements at the scale of entire forests, facilitating large-scale assessments of biomass and carbon storage [75]. Our results demonstrate the potential of topobathymetric LiDAR to facilitate similar investigations in shallow coastal waters, enabling structural analyses and carbon stock estimations for submerged vegetation such as seagrass.

Fine-scale estimation of living biomass and blue carbon stocks of submerged above-ground seagrass meadows in a cost-efficient manner is an inherent challenge to monitoring and mapping actions [76]. Traditionally, it is performed by labour-intensive methods, and insufficient data are often a limiting factor to nature protection agendas and ecosystem accounting programmes [77]. LiDAR point cloud data advance quantitative estimation of biomass and blue carbon from remotely sensed data as a cost-efficient approach, as we here demonstrate a method for a digital and remotely sensed assessment. In this study, the biomass of an entire eelgrass meadow was calculated considering both a precise estimate of the areal extent and the canopy height at a spatial resolution of 10 cm. This is in large contrast to traditional in situ measurements where spatial distribution is often based on point measurements of seagrass presence/absence along predefined transects, with canopy height determined from manual observations using SCUBA or snorkelling at selected locations inside the meadow [78].

Large-scale quantification of seagrass density and biomass has been demonstrated through passive remote-sensing techniques (as opposed to active remote sensing or in situ measurements) via airborne- or satellite-derived products. For example, ref. [20] calculated the leaf area index (LAI), which comes from an in situ-derived mathematical relationship between seagrass leaf area, leaf density per square metre (m⁻²), and seabed reflectance. This relationship can then be validated against, and upscaled to, satellite-derived (e.g., Landsat, Sentinel) seagrass areal extent [20,79]. Yet, such application does not quantify the height (and thus volume) of seagrass beds and also may be limited in resolution, depending on the type of remote-sensing data (e.g., airborne, satellite). Albeit at a smaller spatial scale, our LiDAR application demonstrates areal, high-resolution, volumetric eelgrass canopy measurements. We therefore suggest, for further research, the inclusion of high-resolution, three-dimensional LiDAR data in established seagrass remote-sensing techniques, to facilitate a more robust characterisation of seagrass beds that incorporates seagrass extent, density, and areal canopy height. Aside from canopy volume, topobathymetric LiDAR can also supplement other remote-sensing analyses that utilise elevation as a variable.

In the present study, our LiDAR-derived biomass estimates hold uncertainties, including an areal assessment of volumetric biomass density, plus the relationship between volumetric biomass density and LiDAR point cloud density, as well as the LiDAR-derived canopy height. The biomass density of a seagrass canopy is influenced by the shoot density, leaf length, and weight of individual plants [76]. Here, we estimated the biomass density from the weight and height of the canopy at several locations in the study area (n = 42), and thus assumed an average relationship between leaf length, shoot density, and weight. The shoot density and above-ground biomass have been shown to be correlated, although showing large variabilities, while per-metre biomass becomes independent of shoot density as it approaches maximum biomass [76]. Largely, biomass and cover have been shown to depend on light availability and thus water depth [80,81]. In the present study, all data were sampled at depths shallower than 3 m. Our sampled biomass density had a mean of 323.1 g DW m⁻³ with a standard deviation of 236.8 g DW m⁻³ (Table 2). As such, the uncertainty of the biomass estimate should be considered, whilst the accuracy of the LiDAR-derived canopy height is discussed further below. As it was outside the scope of the present case study, we recommend that further research investigate the mathematical relationship between eelgrass volumetric biomass and corresponding LiDAR point cloud density.

Eelgrass biomass in the Nordic countries exhibits considerable spatial and seasonal variability, influenced by factors such as depth, light availability, and nutrient conditions [76,82]. Our estimated above-ground biomass of the surveyed eelgrass bed (63 g DW m⁻², Table 3) overlaps within a previously reported range of 26 to 546 g DW m⁻², and with values reported in September in the area (~170 g DW m⁻²) [83]. This variability reflects the dynamic nature of these coastal ecosystems and underscores the importance of site-specific data for accurate carbon budget estimates and management actions. Overall, the mean above-ground carbon stock was estimated between 95.9 kg C (50 m) and 122.3 kg C (25 m) at the study site (or between 19.5 and 21.4 g C m⁻², Table 3), corresponding to 352.0 to 448.8 kg CO₂ stored in a small section of the bay. While the aim of this case study was to demonstrate the potential of drone-borne topobathymetric LiDAR, we recommend for further research an updated, site-specific dry weight to carbon ratio, in order to improve fine-scale estimates of the carbon stock.

In addition to biomass and blue carbon content estimation, high-resolution measurements of seagrass canopy height and habitat volume are ecologically relevant, as seagrasses support a range of key ecosystem services [1,84]. Seagrasses modulate hydrodynamic processes by influencing currents, wave propagation, and erosion rates [85,86], thereby providing hydrodynamic shelter for a large number of species living epiphytically on the plants or within the plant canopy [87,88]. Thus, drone-based LiDAR enables new opportunities to link 3D habitat structure to associated fauna and ecosystem services and further advance automated, reproducible monitoring of submerged aquatic ecosystems.

Our results provide a range of estimated eelgrass habitat volumes within the analysed area. By comparing LiDAR-derived canopy heights with ground-truth measurements, we found that in situ values were generally higher, despite being measured after the LiDAR data, when eelgrass canopy height and thickness were reduced [89]. Overall, potential error from in situ canopy height measurements should be considered within this comparison. Nonetheless, this pattern is in line with previous studies in terrestrial grasslands, whereby LiDAR-derived canopy height and estimated above-ground biomass tended to underestimate field observations [90,91]. Reduced LiDAR canopy height estimates may result either from insufficient point returns at the canopy top [90], likely due to the small surface of grass blades being missed by the LiDAR system [92], or from a lack of ground return points at the bottom, as dense canopy cover prevents laser penetration down to the ground [93]. Since our LiDAR data were collected around the seasonal peak of above-ground eelgrass biomass in the region, the raw point cloud revealed little to no sea floor return points under thick patches of eelgrass (see Figure S2). Such dense canopies can therefore create gaps in terrain points, influencing the accuracy of the DTM interpolation [94,95] and thereby influencing calculated canopy height.

Differences in modelled canopy height and sea floor bathymetry between the 50 m and 25 m LiDAR datasets further highlight the influence of canopy loss from both canopy top and/or bottom. With nearly double the point density per square metre, the 25 m LiDAR dataset likely captured more returns both at the canopy top, as well as at the sea floor between eelgrass plants [96], thus explaining the variation in modelled canopy extent, height, and DTMs between the two LiDAR datasets (Figures S4 and S7–S9). Moreover, variation in DTMs could also result from more noise present within the higher-density 25 m point cloud data [97], highlighting this important consideration when operating in areas of low water clarity. These results, along with the influence of the DSM and DTM smoothing windows, highlight the influence of UAV flight elevation, point cloud density, and digital model parameters in determining geometric habitat volume from drone-borne LiDAR, and thus, should be considered in future applications.

Validation of LiDAR elevation values using echosounder bathymetry revealed a vertical offset of approximately 14 cm between the two techniques (RMSE, Table 1), although the linear relationship between the two methods was strong (Figure 6). This offset may reflect methodological differences in how elevation is calculated between LiDAR and sonar, for example, data registration, noise filtering, terrain characterisation, and the distance between the sensors and the sea floor [98,99]. Both LiDAR and sonar have their own advantages and limitations. Sonar can acquire bathymetric data at greater depths and is less affected by water quality, but its accuracy will be limited in shallow waters with complex terrain [100]. Conversely, while drone-borne LiDAR cannot obtain data in deeper waters (as was the case in deeper sections of our study area), it can effectively survey the land–water interface and complex terrain and cover a larger geographic area, as well as derive radiometric information of surveyed surfaces. In our results, we demonstrate the benefits of drone-borne topobathymetric LiDAR as an efficient, high-resolution, and cost-effective way to survey shallow coastal areas and submerged vegetation, which are critical habitats for biodiversity and blue carbon storage.

Furthermore, our use of a commercially available, medium-sized drone with an integrated topobathymetric LiDAR system presents benefits relative to traditional airborne LiDAR bathymetry (ALB) on crewed aircraft surveys. ALB surveys usually rely on sensors fixed to crewed airplanes, which are associated with higher operating costs and reduced reproducibility (i.e., not pre-programmable) compared to drone deployment. Although ALB surveys currently travel faster and cover larger areas versus drones, they generally produce lower point cloud densities. For example, ref. [70] surveyed seagrass beds across ~1100 km² via ALB, achieving an average point cloud density of 4 sea floor points per square metre. While no doubt effective for surveying large areas, ALB resolution may limit structural analysis. In contrast, our drone-borne LiDAR setup, covering a smaller area, produced an average point density of 70 points per square metre of sea floor (50 m dataset), facilitating detailed, high-resolution investigation of marine vegetation structure. The drone-sensor setup mentioned here can cover ~1 km² per day in good flying conditions. Flying the drone at a higher elevation (120 m maximum height, according to the manufacturer), could provide a wider areal coverage, but reduced point density and thus variation in geometric estimates from the point cloud. Our results display the importance of such considerations when mission planning to monitor submerged aquatic vegetation.

In the present study, point cloud classification was limited to two classes (sea floor and vegetation), with high-accuracy classification between the two. Guided by annotations informed via ground-truth data, these results were to be expected, as sandy/rocky bottom will reflect more light than plant surfaces, which tend to absorb more light [101]. Following classification and subsequent digital model creation (DTM and DSM), we manually removed areas containing other types of vegetated habitats. To circumvent this and leverage a computational workflow, we recommend that future applications include multiple vegetation types, based on ground-truth radiometric values specific to individual species or species groups. As well, habitat classifications derived from drone-borne photogrammetry and machine learning (ML) can be used to guide point cloud classification or digital model segmentation [23,24].

One major limitation that must be addressed in topobathymetric LiDAR is related to point cloud noise. In the present study, the majority of noise was cleaned manually. Several algorithms exist to expedite point cloud cleaning, such as statistical outlier removal (S.O.R.), or noise removal via deep learning [102]. We attempted statistical outlier removal, which resulted in many relevant eelgrass points being filtered out, likely related to the complex optical environment across and below the water surface. We therefore relied on manual cleaning, which can be time-consuming, although it is not uncommon in current topobathymetric LiDAR applications [69]. The results from this case study highlight the need for further development of automated point cloud noise removal in topobathymetric LiDAR data, especially for surveys in areas of relatively low water transparency, with complex surface conditions and/or sea floor, and high vegetation cover.

We consider that green LiDAR holds the potential to quantify the lower depth limit of seagrass growth and its relation to bathymetry and environmental gradients, a key indicator of ecosystem status within the European Commission’s Marine Strategy Framework Directive [103,104]. We did not explore the depth penetration relative to light attenuation in this study. However, the green LiDAR instrument used here is considered capable of reaching a depth of ~2 Secchi, which is within the range of the reported lower growth limit for eelgrass, typically, at 1 to 3 Secchi [105]. This, however, remains to be empirically tested, as Secchi depth only represents an approximation of light attenuation, and its measurement is sensitive to the capabilities of the human eye [106]. Drone-borne topobathymetric LiDAR also displays the potential to further map other important submerged habitats and blue carbon ecosystems, including coral reefs [107] mussel or oyster reefs [108], maerl beds, and kelp forests. As well, coastal erosion studies, sediment relocation, and changes in bathymetry and ecosystem structure are areas where topobathymetric LiDAR can improve and contribute through further development for future applications.

5. Conclusions

The potential benefits of drone-borne green LiDAR in environmental and ecological management of shallow coastal habitats are substantial. Our results display the ability of this technology to map areas at high resolution (centimetre-scale), to distinguish between hard bottom and vegetative habitat, and to create 3D models of geometric structure. From these models, we were able to estimate habitat volume and biomass of above-ground living eelgrass tissue, thus facilitating blue carbon content assessments. Considering eelgrass’ shallow-water habitat and patchy distribution, we display the utility and cost-effectiveness of drone-borne LiDAR relative to more traditional, fine-scale methods such as manual in situ monitoring, echosounders, or ALB surveys. This utility also has the potential to advance ecological research and inform management decisions, for example, by providing key parameters such as canopy height and volume and extent. We also demonstrate the potential for high-resolution LiDAR data to complement other seagrass remote-sensing approaches.