Semi-Automated Method for Mapping and Classifying Boreal Coastal Wetland Plant Communities using Drone and Ground Data

Summary

The protocol comprises four integrated phases: Unmanned Aerial Vehicle (UAV) survey, ground-truth data collection, vegetation index computation, and AI-assisted classification and mapping of coastal wetland plant communities. The result is a multilayered geospatial dataset that supports applications in ecology, environmental monitoring, and decision-making for coastal wetland habitat conservation and management.

Abstract

Semi-automated plant community mapping bridges the gap between traditional ecological surveys and modern AI-assisted ecosystem monitoring, enabling site survey and monitoring at scales and speeds previously unattainable. Coastal wetlands, dominated by low-growing herbaceous plants, require a drone-based (Unmanned Aerial Vehicle, UAV) photogrammetric survey to achieve higher spatial resolution and greater flexibility in timing. This protocol consists of four phases, namely UAV-based aerial survey, ground-truth data collection and georeferencing, vegetation index calculation, and supervised classification using the random forest (RF) algorithm in R (i.e., AI‑assisted, semi‑automated mapping of plant community distributions using machine‑learning classification). Key R packages used include terra, sp, sf, rgdal, raster, rsample, MLmetrics, randomForest. The emphasis in the second phase is on a traditional ecological survey method—stratified quadrat ground sampling. Ground sampling serves as ground truthing, providing empirical evidence to ensure the accuracy and reliability of the final map product. Two widely used UAV-mounted sensor types were tested: a multispectral sensor and an RGB camera, yielding 19 multispectral and 27 RGB-based indices, along with an RGB-derived digital surface model (DSM). Plant communities were mapped at a Baltic coastal wetland case study site, and the following algorithm performance results were obtained: using the multispectral dataset, the model achieved an overall accuracy of 92.3% with an out-of-bag (OOB) error of 7.75%; in comparison, the RGB dataset achieved an accuracy exceeding 98% with an OOB error of 1.14%. These results reinforce the suitability of both sensor types. Each phase of the protocol produces georeferenced datasets. These can be compiled into a layered geographical information system (GIS) Project with embedded biophysical field observations, serving as a foundation for complex ecosystem research, such as ecological modeling to forecast impacts and changes. In practice, this GIS Project can serve as a historical record of land cover and support management, environmental restoration planning, and monitoring.

Introduction

The goal of this protocol is to provide a practical, step-by-step workflow that combines drone-based (i.e., unmanned aerial vehicle [UAV]-based) photogrammetric survey with on-site biophysical observations as ground-truth data to produce a coastal wetland plant community distribution map. The final output, a land-cover map, is created through a semi-automated machine-learning process (random forest [RF] algorithm)¹.

A plant community map is an essential dataset in wetland ecological modeling because it transforms complex vegetation patterns into a spatially explicit, pixel-based representation of habitat structure. This baseline dataset enables modeling of ecological processes (e.g., biomass distribution, species interactions, eco-hydrological responses, nutrient cycles) across a landscape^2,3,4.

In addition to the plant community map, following the steps of this protocol produces a collection of geospatial data. Saved as a geographical information system (GIS) project, the dataset can be integrated into a GIS, shared, and used for further ecosystem analysis and decision-making to support management planning and environmental monitoring.

In this method, a UAV survey yields an RGB (i.e., red, green, blue, the three primary colors) and/or a multispectral, geometrically accurate orthomosaic map, along with a digital surface model (DSM), all serving as baseline data^5,6. This data is used to calculate vegetation indices (VIs) using orthomosaic maps as inputs to the GIS software's calculator. To create the final output map, VIs serve as per-pixel predictor variables for RF, transforming the raw spectral information of the orthomosaic into structured inputs for classification^7,8. Because VIs are mathematical combinations of spectral bands that highlight vegetation properties reflected by plants and light⁷, VIs can be used as independent data in ecology to monitor habitat quality and biodiversity, and support ecological modeling^8,9. Therefore, VIs are saved as raster maps and included in a GIS project.

To classify and map plant communities in an AI-assisted environment, remotely sensed datasets (e.g., UAV survey maps, DSM, and VIs) must be combined with georeferenced on-site biophysical observations to ensure ecological reliability, reproducibility, and meaningful interpretation¹⁰. Such field data serve as ground truth for the RF algorithm to classify habitat-vegetation-cover pixel data in VIs and to predict plant community distribution across a study site (e.g., to create a plant community map)¹¹. In addition to serving as a ground-truth dataset for machine learning, georeferenced observations provide a means to anchor biophysical observations within GIS. This way, the ecological dataset functions as an attribute table, which can be used as environmental variables for ecosystem modeling^12,13. On-site data is valuable for evidencing the ecological conditions of a habitat because it is gathered directly within the actual environment under study, providing a realistic and immediate assessment of specific conditions¹⁴.

Coastal wetlands are low-energy habitats located in sheltered shoreline areas. Wetland ecosystems are distributed along the coastlines of six continents, excluding Antarctica, playing crucial ecological roles globally, except for ice-covered landscapes in the extreme polar regions¹⁵. The global population growth and industrialization, coastal defenses, and intensification of agricultural practices significantly impact coastal wetlands, causing fragmentation and degradation^16,17. These pressures make coastal ecosystems a priority for ecological restoration and conservation worldwide¹⁸.

Environmental monitoring and ecosystem modeling of coastal habitats facilitate the evaluation of ecosystem health, the identification of temporal changes, and the assessment of restoration outcomes¹⁹. Nonetheless, effective monitoring can be hindered by the dynamic nature of these ecosystems, their restricted accessibility, and the substantial costs of data collection in complex, often remote intertidal zones. In dynamic, hard‑to‑access habitats, vegetation often provides the clearest, most stable signal of ecological change¹⁸.

As vegetation is a fundamental basis for habitat formation, classifying plant communities and mapping their distributions provides a critical foundation for ecological assessment and habitat process modeling^20,21,22. Coastal wetlands are highly heterogeneous systems, exhibiting significant variation over short distances in elevation, salinity, vegetation structure, hydrology, and sediment characteristics. Due to this complexity, the protocol incorporates stratified quadrat sampling²³ (minimum ten 1 m × 1 m quadrat samples per plant community), which is not only beneficial but methodologically imperative for generating representative ecosystem models and land cover maps^24,25.

To illustrate this methodology, a case study was conducted on a microtidal marsh on Hiiumaa Island in the West Estonian archipelago of the Baltic Sea (Figure 1). The Baltic Sea's coastal landscape, including marshes in sediment-accumulation zones across more than 100,000 islands within the archipelagos, is managed at different levels of intensity²⁶. For biodiversity conservation and restoration, Baltic boreal coastal wetlands are maintained as semi-natural grasslands^27,28. Such wet grasslands have regional landscape heritage value, and some, including the case study site, are part of the Natura 2000 network of protected areas. In such areas, to sustain low-growing vegetation, which supports greater biodiversity and provides habitats for wildlife such as waders²⁹, coastal meadows are grazed extensively and/or mown seasonally to maintain open grasslands and prevent them from being overtaken by expansive species, for example, Phragmites australis³⁰. In boreal coastal regions, wetlands managed under conservation measures typically include four key plant communities: Open Pioneer (OP), Lower Shore (LS), Upper Shore (US), and Tall Grass (TG)³¹. All four occur at the selected case study site. Vascular plant species dominate coastal wet grasslands. The OP indicator species are halophytes: Salicornia europaea (most common height range 3–30 cm) and Suaeda maritima (10–50 cm). The LS is dominated by Glaux maritima (3–25 cm) and Juncus geradii (25–75 cm). US indicator species are Festuca rubra (15–90 cm) and Leontodon autumnalis (10–80 cm), and dominant grasses identify TG: Deschampsia cespitosa (30–120 cm), Elytrigia repens (30–120 cm), and Molinia caerulea (50–180 cm). In addition, typical coastal wetland plants in this climate zone include: Spergularia marina (5–30 cm), Plantago maritima (10–30 cm), Triglochin maritima (15–70 cm), and Agrostis stolonifera (15–30 cm), which occur across communities.

Coastal wetlands are defined by their low-relief landscapes with subtle microtopographic variations, such as depressions and hummocks, that influence plant community distribution. These landscape features form naturally through processes such as sediment deposition during floods, storm surges, and tidal currents³². These features often lack distinct signatures detectable by coarse aerial imagery, but species composition reflects specific spectral responses that can be recorded and classified³³. Commonly, satellite and aircraft aerial mapping provides the geospatial data needed for much ecosystem surveying and modeling to support monitoring and assessment^34,35. However, low-altitude drone-based aerial mapping is essential for effectively surveying coastal grasslands. This approach delivers high-resolution data and provides a more detailed photogrammetric record of landscape features, including vegetation distribution patterns and vegetation height models derived from a Digital Surface Model (DSM), thereby improving classification accuracy³.

Compared to alternative approaches, UAV workflows allow flexible scheduling to match phenological stages, minimize cloud interference, and capture structural detail unavailable from satellites^36,37. Combined with machine learning (e.g., Random Forest), UAV data support automated pattern detection, object classification, and the extraction of valuable insights from complex datasets that would be overwhelming for humans³⁸. However, on-site ground sampling remains indispensable for supervised learning because it provides verified labels for training and assessing models, ensuring that models learn correct patterns and that their performance is robust and reliable^11,39,40.

This protocol is designed for low-relief ecosystems with low-growing plant communities that are accessible for on-site sampling. The workflow is divided into four distinct phases (Figure 2): (1) UAV-based aerial survey, (2) ground-truth data collection and georeferencing, (3) vegetation index calculations, and (4) supervised classification using Random Forest in the open-source statistical computing environment R. The protocol was implemented using RStudio with the following key packages: terra, sp, sf, rgdal, raster, rsample, MLmetrics, and randomForest.

Protocol

NOTE: The following steps describe a detailed workflow (see diagram in Figure 3) for mapping and classifying coastal wetland plant communities. This protocol can be replicated using either a UAV-mounted multispectral sensor or an RGB camera. Completing this protocol will require several days. Therefore, the methodology is divided into four phases, with optional intermissions. The prescribed sequence is recommended as outcomes of the previous phase support the execution of the next. The first three phases generate datasets that serve as inputs to the final phase.

1. UAV-based aerial survey

1. Acquire and review essential precautions for UAV operators, local regulations, and the manuals for drone hardware and software.
   1. Ensure the pilot operates safely, legally, and professionally; hold all required certifications or licenses for the region of operation; and be fully trained on the specific drone model and mounted sensor/camera use and mission workflow. Consult the state or regional regulatory framework for acquiring UAV insurance.
   2. Identify the airspace classification for the mission area and obtain required authorizations. Respect altitude limits, no‑fly zones, and proximity restrictions around airports, infrastructure, and populated areas. For environmental and wildlife considerations, follow seasonal restrictions (e.g., nesting periods) and comply with environmental regulations for protected landscapes, wetlands, Natura 2000 sites, and other protected areas.
   3. Notify local community members or stakeholders, where necessary, to secure approvals and prevent misunderstandings. Clarify data protection rules for when aerial images may capture people, private property, or vehicles by ensuring everyone understands that the drone flight is intended to map plant communities for research, and that any additional data subject to data protection laws will remain confidential.
      NOTE: Notification is typically necessary when national or regional regulations mandate prior communication for drone operations.
   4. Follow all manufacturer guidelines for assembly, maintenance, and operation of the drone and mounted sensors; such compliance ensures safe operation and protects equipment integrity. Respect the maximum wind limits (9 m/s, including gusts), temperature ranges (Min. 0 °C; Max. 35 °C), and payload capacities when selecting sensors/cameras. Store and transport batteries in accordance with the manufacturer's instructions and aviation safety regulations.
2. Design a UAV flight mission.
   ​NOTE: To implement this protocol, the eBee X fixed-wing UAV system was used, equipped with the multispectral and RGB sensor and the flight management software. Always consult the manufacturer's manual when using a UAV.
   1. Open flight management software preinstalled on a laptop. On the welcome page, under the New Mission tab, select the camera to use (in this case: Sequoia), enter a descriptive mission title (e.g., the study‑site name), and select Create mission. The Mission panel will open automatically, and the satellite map will appear in the main view.
   2. On the satellite map, use the cursor to zoom, pan, and navigate to the approximate location of the study site. In the Working area tab, select Place working area and position the cursor near the center of the study site, and click. A circle will appear representing the working area; adjust its size by increasing or decreasing the radius.
   3. On the Mission panel, click the Take-off and landing tab. Then click on Add New Start, navigate the cursor on the satellite map to the desired take-off location, and click to set a new starting point. After that, click on Add new Home, choose Linear landing, navigate the cursor on the map to the desired UAV landing/return point, and click to set.
      NOTE: When planning missions, use the software's satellite map and measurement tools to find optimal landing and take-off areas, as an obstacle-free landing strip is required. Some fixed-wing UAVs do not have vertical take-off and require a gradual climb to altitude during take-off and a gradual descent. Such UAVs need a landing strip measuring 200–300 m long and 75 m wide.
   4. On the Mission panel, select the Mission blocks tab. Click on Add mission block, then select Horizontal mapping. From the Camera drop-down menu, select Sequoia, and from the Plan above, select Elevation data - EAD. The presettings menu will appear with tabs with adjustable values: Resolution 10 cm/px, Lateral overlap 75%, and Longitudinal overlap 80%, resulting in a flight altitude of 106.1 m/AED (these automated values were selected for the UAV survey).
   5. Place the cursor on the satellite map view and draw a polygon around the study area by left-clicking to add adjustable corner points around the survey site. To finalize, right-click to display the potential drone flight path on the map.
      ​NOTE: To implement this protocol, the survey area was estimated to cover 40 ha with a flight time of 32 min.
   6. Still in the Mission panel, in the Safety actions tab, ensure the following actions are selected. For Return to Home: if strong wind is detected, if the camera reports a fault, in low light (or lens cap is on), if endurance is low, if GNSS accuracy degrades, if ground modem link is lost for 300 s (can be adjusted), and for the Climb: if ground proximity is detected, should be selected.
   7. Timing is critical: Schedule the mapping mission as close as feasible to the ground-truthing (biophysical sampling date). This is vital, as it guarantees that drone imagery and field data reflect the ecosystem under the same environmental and phenological conditions, significantly enhancing the scientific precision and interpretability of maps.
   8. Select a low-tide window (or “low-water” conditions in non-tidal coasts) to maximize exposed habitat and minimize water glare; when timing aerial surveys, also consider wildlife disturbance and try to avoid breeding seasons.
3. Perform the pre-flight checks and flight mission.
   1. On-site, assess environmental conditions. Verify current sea water levels are low; avoid disturbing wildlife; ensure the take-off and landing area is stable, dry, and sufficiently elevated above water. Examine the airspace region where the flight mission will occur for any aerial obstacles, including the take-off and landing areas.
   2. Monitor the weather forecast for optimal flight conditions: wind (max 9 m/s), visibility (no fog), humidity (below 75%), and precipitation (max prediction 40%).
   3. Assemble UAV system: drone hardware, sensor/camera, antenna, and flight management software. Connect a drone battery. Check the UAV hardware, including sensor/camera health; ensure the lens is clean; and verify the sensor/camera is correctly connected to the UAV.
   4. Switch on the fieldwork computer, open the flight management software (eMotion), and click Load Mission from the Browse file.
      1. Connect the USB antenna to the laptop, click Connect, and a flight control bar will appear on the right.
      2. In the Camera tab, confirm that there is sufficient memory available to store all survey images (empty SD card, min 32 GB), and that the correct camera (Sequoia) is listed. Enable the computer's sound to hear software notifications.
   5. Verify that firmware and software are updated (via automated software notifications) and calibrate the compass if needed.
   6. Ensure the predesigned mission is correctly loaded (i.e., flight path, take-off, and landing trajectories are visible in the main view). Ensure safety actions are correctly selected (as listed in step 1.2.6).
   7. Calibrate the multispectral sensor. Place the panel on the ground, then hold the drone over it with the sensor/camera lens facing the panel.
      1. In software, access the flight control bar, open the Camera tab, and press Calibrate.
      2. Wait for sound signals to indicate picture-taking until the text Radiometric calibration complete appears. This is not needed if using only the RGB sensor.
   8. Before starting the mission, check the wind direction. Grab a small handful of light, dry grass, sand, or dry leaves. Release the material from eye level; the direction the grass drifts indicates the wind's path.
   9. Then, in the software's main view (satellite map), adjust the UAV flight path orientation by dragging the orientation handle so that the flight lines run perpendicular to the wind direction. Ensure the flight path covers the entire survey site: adjust polygon corners by dragging corner points.
   10. Adjust landing trajectory orientation to ensure landing path orientation is against the wind.
   11. Ensure the flight path, take-off, and landing areas are within the working area after all the adjustments. Expand the working area circle, if necessary, by increasing the radius numerical values in the Working area tab.
   12. Start the UAV flight as an automated flight mission following the predesigned flight path (survey area for the case study: 40.5 ha). Log flight starting time.
       1. During the flight, keep the flight management software on and the drone antenna always connected.
       2. Listen to and observe the software for battery-life updates, sensor performance, and the drone's alarm signals during the mission (e.g., wind-speed increase, obstacle detection).
       3. Maintain a visual line of sight (VLOS) with the drone on a mission at all times, unless otherwise approved. Avoid all distractions and stay ready to respond to notifications or perform an emergency landing. Log flight end time.
          ​NOTE: Best practice (depending on region, it can be a legal requirement) is to conduct surveys with two operators: one maintaining VLOS and the other monitoring the flight management software. In the event of a UAV accident, suspend flight operations, secure the area, document the incident, notify relevant authorities, and initiate insurance procedures.
4. Perform post-flight procedures.
   1. Extract all aerial images from the UAV-mounted sensor or camera and back up all survey data, including the flight log/s, onto a computer and/or cloud.
      NOTE: The risk of data loss increases over time, as equipment can be damaged and SD cards can become corrupted.
   2. Verify image sharpness, exposure, and coverage, checking for gaps in the image sequence which may have occurred due to sensor error or caused by glare, shadowing, or tidal inundation. Reschedule and resurvey as needed, ensuring the drone flight occurs in close temporal proximity to the ground-truthing.
   3. Download and save the receiver independent exchange format (RINEX) file from public archives such as national land survey services, corresponding to the flight date and time.
   4. On an external hard drive (250 GB-1 TB), create a folder for the study case (suggested title: date_ location [Avoid using spaces and special characters]). This will be the master folder for data storage and access for subsequent steps, and the location where the geo-package layers will be created/saved.
   5. To perform aerial survey image geotagging, in the flight management software, open the Postflight tab. If the same computer was used to fly in the Calendar, select the flight date. If not, select No.
   6. Browse and upload the flight log from the backup file. Create a project folder on the external memory drive for the study case (title project as suggested on the screen). Click Next.
   7. Browse and upload flight logs from the SD card used during the flight, which can be identified by the flight number, or upload from a backup file. Wait until the flight path appears in the main view of the software. Verify that it is the correct area/flight log. Click Next.
   8. From the drop-down menu under Post processing options, select PPK: Process raw drone's GNSS data. Under Base station log select Rinex, browse for and find the saved RINEX file. Click Next, then wait for calculations to finish. Ensure PPK fixed > Geotags post processed appear 100%, then click Next.
   9. In the Load images drop-down menu, choose one of the following options: from SD card or Browse to find the folder with the aerial survey images corresponding to the flight log. Click Next. Wait until Image matching is successful appears, and click Next to start the geotagged image import into the project folder. These images will serve as inputs in a photogrammetry process.
      NOTE: For each flight log, image geotagging must be performed separately (repeat steps 1.4.4 –1.4.8). For example, after a UAV battery change, the subsequent aerial survey will have a new flight log.
5. Process photogrammetric data (i.e., create map products using aerial survey images).
   1. On the workstation, connect the external hard drive containing the geotagged UAV survey images and open the preinstalled photogrammetry software.
      1. Create a new project and save it to the external drive (recommended format: prefix MS or RGB + location + date).
      2. Add the images, noting that the multispectral and RGB datasets must be processed in separate projects; hence, use the MS prefix for multispectral imagery and RGB for the latter.
         ​NOTE: PIX4Dmapper was used; however, alternative photogrammetry software can be used to reproduce this protocol.
   2. Ensure that in the Image Properties pop-up screen, the following are listed/selected: WGS 84 as the coordinate system, the correct number of Geolocated images, and the Selected Camera model listing bands (Sequoia and Green, Red, Red edge, NIR bands, or RGB).
      1. Press Next, then choose the processing template: Ag multispectral for multispectral, but RGB Ag for RGB images. Finish, and a satellite map with image locations will appear in the main view.
         NOTE: For the multispectral project, make sure to add calibration images with a calibration panel visible (created in step 1.3.7 and saved automatically with all the survey data). For this case study, a total of 2,320 multispectral and 577 RGB images were geolocated.
   3. Open the Processing options window, click 1. Initial processing here under Key Points Image Scale, select Full and Generate Orthomosaic Preview in Quality Report, then click 2. Point Cloud and Mesh > Point Cloud in this tab, select 1/2 (Half image size, Default) and select Classify point cloud.
   4. Click 3. DSM, Orthomosaic, and Index > Index calculator tab select GeoTIFF and Merge Tiles and all available Indices to be generated (Figure 4).
      1. For the multispectral project, ensure, under Radiometric Processing and Calibration, that each band's Calibration type is set to Camera, Sun Irradiance, and Sun Angle.
      2. Click DSM and Orthomosaic under Raster DSM select GeoTIFF and Merge tiles, plus under Orthomosaic select GeoTIFF and Merge tiles. Click OK and start processing.
         NOTE: When processing multispectral images, calibration type is automatically set using calibration images taken on-site before the flight (step 1.3.7), but can be recalibrated manually by clicking Calibrate. Then browse to find an image of the calibration panel (with the corresponding band title) and input the reflectance factor value read from the calibration panel.
   5. Photogrammetry processing concludes with generated/saved reflectance maps (UAV survey maps) in the multispectral project folder: green, red, red edge, NIR, and in RGB: red, green, blue, grayscale, orthomosaic, and DSM (digital surface model). Ensure that all indices, orthomosaic, and DSM GeoTIFFs are merged and verify the project reports for errors; if issues are detected, repeat the processing.

2. Collecting ground truth data and performing georeferencing

1. Prepare for fieldwork.
   1. Visit the study site in advance and explore the area to identify potential plant communities of interest, thereby gaining experiential knowledge of the site. Decide which biophysical observations are feasible to collect on-site and are necessary for describing the plant community composition of the selected study site.
      NOTE: To illustrate this protocol, the focus was on vegetation species, structure, and composition to classify 4 plant communities indicative of a Baltic Boreal coastal meadow habitat: Open Pioneer (OP), Lower Shore (LS), Upper Shore (US), and Tall Grass (TG).
   2. Prepare and print a visual plant species identification guide listing the species indicative of plant communities in the particular habitat(s) being mapped. The guide can speed up and assist with on-site species identification if electronic devices fail.
   3. Prepare a worksheet for recording field observations, including a list of indicator species along with other species commonly found in the habitat(s) under study and the surrounding area, as well as additional biophysical field observations that need to be recorded (e.g., survey date, vegetation heights, biomass weights).
2. Use a stratified sampling procedure to collect ground truth data using quadrats.
   1. Identify a plant community visually based on experiential field knowledge.
   2. Place a quadrat frame on the ground at a representative location for the plant community being sampled, avoiding areas with significant disturbance. For Boreal coastal wetlands, a quadrat of 1 m x 1 m is sufficient. This quadrat will serve as a sample plot for biophysical observations and biomass sampling.
   3. Record the survey date, surveyors' names, plant community type/class, and the number code for each sample quadrat.
      NOTE: It is recommended to establish number codes in the following order: study field code + plant community code + sample quadrat sequence number of the quadrat being recorded. These number codes will link biophysical field observations to the geospatial dataset and coordinates.
   4. Record coordinates on each corner of the sample quadrat using GNSS/RTK systems and land surveying software, by inputting a number code corresponding to the quadrat number code and saving the coordinates.
      NOTE: Use a GNSS receiver that provides high-precision (centimeter-level) location data; refer to the user manual for instructions on operating the device and its complementary software.
   5. As for biophysical observations, within the sample quadrat, record estimated cover percentages for each observed plant species, as well as the cover of other potentially indicative variables, including bare ground, plant litter, moss, and algae (ensuring that the total is between 95% and 105%, an accepted range for phytosociological surveys). Take five random measurements of the vegetation height using a measuring tape to determine the average sample height.
   6. With the same sample quadrat, collect optional biophysical observations: for example, soil moisture percentage, sample vegetation biomass by cutting vegetation at the soil level, collecting biomass into a paper bag, and weighing using a battery-operated, leveled scale (precision ± 1 g). To find the true biomass weight, subtract the paper bag's weight from the total weight. Record all weights for future reference.
3. Replicate the stratified quadrat sampling procedures described in previous steps (2.2.1–2.2.6) a minimum of 10 times within each of the plant community classes (i.e., OP, US, LS, TG).
   ​NOTE: Avoid placing sample quadrats too close together. It is essential to provide consistent, comprehensive sampling across the study sites to ensure that the recorded quadrat samples accurately reflect the habitats. Since plant communities often occur in mosaic patterns in coastal wetlands, sample the same plant community at different locations within the study site.
4. Digitalize quadrat sampling records after completing all field observations.
   1. Transcribe all fieldwork notes into spreadsheet software, creating a dataset for analysis by organizing the data in a table format with a header row labeling each variable; then input all the values accordingly.
   2. Save the file and export it as a comma-separated values (CSV) file.
5. Create a ground truth dataset by georeferencing biophysical observations.
   1. Download all the recorded coordinates of the sample quadrat corner locations as a CSV file from the GNSS/RTK system used in the field.
   2. Create a new project in Geographic Information System (GIS) software (QGIS was used to implement this protocol), and save it as a .qgz file into the project created in step 1.4.3. This folder and the .qgz file will include all geospatial datasets from the forthcoming steps.
   3. In the GIS workspace, import a CSV file containing the coordinates of quadrat corners as a Delimited Text Layer. After importing these coordinates, they will appear as shape: points.
   4. Join the coordinates of each quadrat corner point as a shapefile: polygon, and save the shapefile.
   5. Join the polygon shapefile with the digitized field dataset CSV file (created in step 2.4) by linking the number codes of both files.
   6. Save the newly created ground truth dataset as a GeoPackage (.gpkg) file in the folder created on the external hard drive (step 1.4.4). Save to update the GIS Project.

3. Calculating the vegetation index (VIs)

1. Prepare UAV-based survey maps for the Vis.
   1. Into the GIS software workspace (use the same .qgz file created in step 2.5.2 and updated in 2.5.6), import all reflectance mosaics (map products) created using the UAV survey at the first phase of the protocol, in step 1.5.
   2. Reduce the size of the maps by 'clipping' unnecessary data. Looking at the sample quadrat distribution and based on experiential knowledge of the study site's extent, create a polygon boundary surrounding the study site and clip maps to the polygon's extent (Figure 5). This will reduce the file size and increase processing speed.
      ​NOTE: Clipped survey maps (multispectral GREEN, RED, Red edge, NIR, and RGB red, green, blue band GeoTIFFs and ground_truth.gpkg (created in step 2.5.6) are available to download from the Zenodo repository via link: https://doi.org/10.5281/zenodo.20075625 for educational purposes only to practice reproducing the following steps.
2. Use the Raster Calculator tool to compute; the VI formulas are listed in Supplementary File 1^{34,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71}.
   1. When calculating multispectral VIs, use RED, GREEN, NIR, and Red-edge GeoTIFFs as input values. (see Figure 6, GIS Raster Calculator, one VI formula input as an example).
   2. For RGB-based VI computations, first normalize the red, green, and blue indices (map products from step 1.5). Calculate new RGB VIs only after completing the normalization process (formulas for normalizing RGB indices are available in Supplementary File 1 as part of the RGB VI list).
   3. Save new VIs as GeoTIFFs (e.g., raster files) in separate folders, naming them accordingly (e.g., prefix (MS or RGB) + index name) in the GIS Project folder created on the external hard drive (step 1.4.4).
   4. Add the DSM and other automated VIs (e.g., grayscale) created/clipped in step 3.1.2 to the RGB VIs folder. These folders containing VIs, coupled with a ground truth file (created in step 2.5.6), serve as baseline datasets for a Random Forest algorithm to classify map plant communities. Save to update the GIS Project with links to VI raster files.

4. Supervised classification in R using Random Forest

NOTE: To execute this phase, open a statistical computing environment for R in RStudio, then go to File > Open File and select the script provided in Supplementary File 2. This script is ready-to-use and organized to ensure reproducibility and facilitate adaptation to other study sites. It can be executed in two ways: (1) Run the entire script at once: Click Source or press Ctrl + Shift + Enter, making sure the folder and GeoPackage file paths are updated; or (2) Run Line by Line or Block by Block: Place the cursor on a line or select a block and press Ctrl + Enter. This option is recommended since it offers better troubleshooting alternatives.

1. Prepare R environment.
   1. Install and load required packages (i.e., terra, sp, sf, raster, rsample, MLmetrics, randomForest, rgdal)⁷². 
      NOTE: If rgdal is incompatible with the newest versions of R, the functions stack() and readOGR() can be replaced with rast() and st_read(), respectively.
   2. Set the working directory.
2. Perform raster Preprocessing.
   1. Load all raster files from the surveyed area located in a specific working directory.
   2. Ensure consistent projection, resolution, and extent among the input rasters. If inconsistencies are detected, use the function resample() in R (e.g., terra::resample() or raster::resample()) to align the rasters to a common reference file (typically the one with the coarsest spatial resolution to avoid artificial upscaling)(see article by Hijmans et al.⁷³).
3. Perform data extraction:
   1. Import ground truth data from a GeoPackage (created in step 2.5.6).
   2. Use the terra:: extract() function to link raster values to sample polygons.
   3. Join the data frame containing the extracted values with plant community codes, ensuring they are factors.
4. Split the dataset into training and validation sets using stratified sampling to maintain class balance. Modify the ratio between training and validation samples as preferred, or depending on the availability of samples.
   NOTE: Smaller datasets typically require an 80/20 split to provide sufficient training data. Larger datasets: 70/30 is sufficient and gives a stronger validation or test set (see the publication by Mohammed and Alsunosi⁷⁴).
5. Perform model tuning.
   1. Define a parameter grid to explore different combinations of mtry and ntree.
      NOTE: mtry is the number of variables (predictors) that are randomly sampled as candidates for each split when growing a decision tree in RF. The default value is typically the square root of the total number of predictors for classification and one-third for regression. Whereas ntree is the total number of decision trees to grow in the forest. A higher number of trees generally improves performance but increases computational time; the error rate typically stabilizes after a certain number of trees.
   2. Loop through the grid (i.e., grid search, which systematically tests different combinations of hyperparameter values to find the best-performing set for the model) and evaluate accuracy for each combination.
   3. Choose the combination with the highest accuracy.
      ​NOTE: See Belgiu and Drăguţ⁷⁵, also Bergstra et al.⁷⁶, for a better understanding of the relevance and alternative approaches for this step.
6. Fit the RF model using the best parameter combination. The script records the Out-of-Bag (OOB) error and variable importance.
7. Perform prediction and accuracy assessment.
   1. Predict the plant community class for all pixels in the raster stack and validation dataset. Export the classification map as GeoTIFFs (see Figure 7, which shows a R console screenshot including plot visualizations of the final outputs; Figure 7A shows the plant community map product when using a multispectral VIs input dataset, and Figure 7B shows the plant community map product when using an RGB+DSM dataset).
   2. Compute the confusion matrix, accuracy, F-1 score, precision, and recall on the validation set (see Figure 8, which shows the R screenshots demonstrating VI performance rankings; Figure 8A shows a screenshot of the R console and plots when using multispectral VIs as the input dataset, and Figure 8B shows a screenshot when using an RGB+DSM dataset; Figure 8A, B views include variable-importance plots named "final_rf"). For more details on these metrics, refer to the article by Naidu et al.⁷⁷.
8. Reopen QGIS project (created in step 2.5.2 and updated 2.5.6 and 3.2.3), add plant community classification maps (GeoTIFFs). Save the QGIS project to update the GIS Project.
   1. To share and back up all the geospatial datasets (UAV survey maps, ground truth, VIs, plant community classification maps as a GIS Project), go to the folder for the study site, right click, and select Send to > Compressed (zip) folder.
      NOTE: The complete GIS Project folder may constitute a substantial dataset (this 40 ha case study dataset exceeded 140 GB). If necessary, communicate only the final results and ground truth datasets to reduce file size.

Results

The semi-automated method for mapping and classifying plant communities is designed to generate accurate land-cover information for coastal wetlands, which are dominated by heterogeneous, low-growing herbaceous plants and characterized by microtopography; see an example of such a habitat in the photograph in Figure 1. A case study of the Baltic boreal coastal wetland on Hiiumaa Island, Estonia, was conducted. On-site, four plant communities were sampled for training the classification model: Open Pioneer (OP), Lower Shore (LS), Upper Shore (US), and Tall Grass (TG). Figure 2 provides a concise overview of the methodology implemented in four distinct phases. In contrast, Figure 3 illustrates a detailed diagrammatic guide delineating numbered, actionable steps coupled by arrows to represent the data flow, thus synthesizing the methodology across the steps and phases of the protocol. The final outputs of this method, utilizing a multispectral sensor and an RGB sensor, are multispectral and RGB-based plant community maps, which are stored as layers in a GIS project together with all intermediate datasets. The quality of the geospatial datasets was evaluated upon the completion of each protocol phase, as the quality of outcomes in succeeding phases depends on the quality of preceding-phase results, culminating in the final phase, where inputs comprise the aggregate of all prior steps.

The following summarizes the outcomes of each of the four phases of the protocol:

In the first phase, the UAV aerial survey was conducted at an altitude of 120 m, yielding a Ground Sample Distance (GSD) of approximately 10 cm/pixel. The aerial survey produced multispectral and RGB images. After post-processing kinematic (PPK), geotags were corrected for each drone image, and the images were mosaicked in professional photogrammetry software to produce aerial survey maps. Photogrammetric processing of multispectral aerial imagery yielded raster files of RED, GREEN, NIR, and Red-edge reflectance.

Multispectral and RGB aerial images were processed as separate projects due to the inherently distinct resolutions of photographs. See Figure 4 of the RGB image mosaicking settings. Figure 4A. processing options by selecting merged tiles and generating seamless, unified output maps. Figure 4B. selection of automated index calculations from RGB camera images produced four raster maps: one per band (red, green, and blue), and the DSM (Digital Surface Model). Figure 4C. demonstrates the generation of automatic tie points to match common pixels across overlapping images during the initial processing step. Together with 3 RGB (red, green, blue) indices and DSM, the photogrammetry software generated a software-generated grayscale raster: a monochromatic, single-channel representation that scales and combines the different bands (e.g., red, green, blue), with low pixel values appearing black and high values appearing white. The formulation of this grayscale index is software‑specific and not standardized across photogrammetry platforms.

Photogrammetry software generated quality reports describing the processing results for both sets of aerial image mosaicking projects: multispectral and RGB. These reports provide key metrics to validate accuracy. Indicators of high-quality results are low reprojection error: the closer to 0, the better (anything below 1.0 is generally accepted). For example, dense "match lines" mean strong matches between most images. The output maps were visually assessed and found to be free of voids and other visual discrepancies. Visual discrepancies are a sign of possible reflectance and mosaicking errors, as well as underlying georeferencing grid issues. The photogrammetry report, along with specific software manuals and training resources, is used to identify the issue, find a solution, and reprocess aerial images.

UAV aerial survey maps served as a baseline dataset for recalculating complex spectral information into standardized pixel‑based indices (i.e., vegetation indices VIs generated in the third phase), enabling their use as structured inputs for machine‑learning classification (in the final phase of the protocol).

On-site biophysical observations were collected as a ground-truth dataset. At the case study site (Baltic Boreal coastal wetland), four key vascular plant communities were identified and sampled: Open Pioneer (OP), Lower Shore (LS), Upper Shore (US), and Tall Grass (TG). Utilizing the stratified quadrat sampling method delineated in the second phase of the protocol, the OP plant community was identified by the presence of indicator species: Salicornia europaea and Suaeda maritima; LS by high percentages of Glaux maritima and Juncus geradii; US by its indicator species: Festuca rubra and Leontodon autumnalis; while TG is distinguished by a significant presence of tall grasses: Deschampsia cespitosa, Elytrigia repens, and Molinia caerulea. At least 10 quadrat samples were collected for each plant community. Each quadrat sample included the following biophysical observations: plant species composition (%), soil moisture (%), vegetation heights in centimeters, and precise quadrat coordinate locations. These variables were digitized as a tabulated dataset, imported into GIS Project, added to the sample quadrat GNSS data shapefile, and saved as a GeoPackage (.gpkg). Such a platform-independent file, containing biophysical observations of wetland plant communities and precise sample coordinate data, serves a dual purpose: for further environmental and ecological ecosystem assessment and modeling. Most importantly, for the successful execution of this protocol, this GeoPackage file serves as model training and validation data in the machine learning process delineated in phase 4. Plant community coding: OP, LS, US, TG; were used as training labels in the machine learning process. Accuracy (High-fidelity labels, human audit, consistent format) and structure (clean, rectangular format (rows/columns)) are crucial for producing a high-quality ground truth dataset. Ground truth labels must have near-zero error rates (e.g., typos) to ensure consistent labeling conventions. For example, if "OP" is the ground truth, ensuring it is not labeled in any other way (such as "op", "O.P.”) is essential to avoid duplication.

To execute the third phase of the protocol, the datasets from the aerial survey conducted in the first phase were used to recalculate raw aerial survey data into Vegetation indices (VIs). First, to reduce dataset size and streamline the computing process, photogrammetry outputs "clipped" to omit irrelevant or low-quality pixel data outside the boundaries of the study site (see Figure 5 as an illustration of actions performed in a GIS software).

Then, 19 multispectral VIs were calculated using RED, GREEN, NIR, and Red-edge reflectance maps. The following multispectral VI raster files were generated and included in the GIS Project: Datt index 4 (Datt4)⁴¹, Enhanced Vegetation Index (EVI)⁴², Green Chlorophyll Index (GCI)⁴³, Greenness Difference Index (GDI)⁴⁴, Generalized Difference Vegetation Indices (GDVI)⁴⁵, Green Infrared Percentage Vegetation Index (GIPVI)⁴⁶, Green Normalized Difference Vegetation Index (GNDVI)⁴⁷, Green-Red Difference Index (GRDI)⁴⁴, Green-Red Vegetation Index (GRVI)⁴⁵, Modified Normalized Difference Vegetation Index (mNDVI)⁴⁸, Modified Soil-Adjusted Vegetation Index⁴⁹ (MSAVI: see Figure 6. a screenshot of a raster calculator in GIS software with MSAVI formula input), Modified Simple Ratio red-edge (MSRred_edge)⁵⁰, Normalized Difference Water Index (NDWI)⁵¹, Normalized Difference Vegetation Index (NDVI)⁵², Red-Edge NDVI (NDVIre)⁴⁷, Red-Edge Triangulated Vegetation Index (RTVIcore)⁵⁰, Soil Adjusted Vegetation Index (SAVI)^53,54, Simple Ratio (SR)⁵⁵, Red-Edge Simple Ratio (SRre)⁴⁷.

Furthermore, a total of 23 RGB-based VIs were calculated and saved as raster files in the GIS project for further spatial analysis: Blue–Green Ratio Index (BGRI)⁵⁷, Brightness Index (BI)⁵⁸, Blue Wide Dynamic Range Vegetation Index (BRVI)⁵⁹, Color Index of Vegetation (CIVE)⁶⁰, Enhanced Green View Vegetation Index (EGVI)⁶¹, Enhanced Vegetation Index (ERVI)⁶², Excess Green indices ExG and ExGI⁶³, Combination index (COM)⁶⁴, Green Chromatic Coordinate (GCC)^34,65, Green Leaf Index (GLI)⁶⁶, Green-Red Vegetation Index (GRVI)⁶⁷, Modified Green Red Vegetation Index (MGRVI)⁶⁸, Normalized Green-Blue Difference Index (NGBDI)⁶⁹, Normalized Green-Red Difference Index (NGRDI)⁶⁹, Red–Green–Blue Ratio Index and Red-Green-Blue Vegetation Index (RGBRI⁷⁰ and RGBVI⁶⁸), Red-Green Ratio Index (RGRI)⁵⁷, Visible Soil Adjusted Vegetation Index (SAVI)⁵³, Triangular Greenness Index (TGI)⁶⁹, Visible Atmospherically Resistant Index (VARI)⁴⁷, Vegetative Index (VEG)⁷¹, Woebbecke Index (WI)⁶³.

Final outputs include two separate multispectral and RGB plant community classification maps saved as GeoTIFFs within the GIS project, along with all the preceding geospatial data. 19 Multispectral VIs, along with the ground-truth dataset, were used as pixel-level predictors in a machine-learning workflow (e.g., RF algorithm) to classify Boreal coastal plant communities (OP, LS, US, TG) and generate a multispectral plant community map. And RGB-based predictors (25 RGB VIs, grayscale index, and normalized⁵⁶ red, blue, green indices and DSM), along with the ground-truth dataset, to classify and generate an RGB-based plant community map (see Figure 7. Final outputs plotted in the R console; Figure 7A represents multispectral, and Figure 7B represents RGB-based plant community maps).

To generate final outputs, the R environment (phase 4 of the protocol) was used to script and execute the machine learning workflow (e.g., RF algorithm), enabling visualization of intermediate outputs, such as vegetation index (VI) performance rankings (Figure 8). When mapping and classifying plant communities, the R console provided real-time error messages and explanations, facilitating troubleshooting and ensuring smooth progression through each step.

This study evaluated two RF algorithm models for plant community classification maps. The first RF model, using multispectral VIs as the input dataset, achieved 92.34% accuracy with a F1 score of 0.915 in the validation dataset. This is a valid outcome, as F1 scores range from 0 (worst) to 1 (perfect), with higher scores indicating superior performance. The F1 score in an RF model represents the harmonic mean of precision and recall. This metric effectively balances false positives and negatives, i.e., it does not allow one type of error to outweigh the other, which is essential for imbalanced datasets, and is frequently averaged across classes or reported per class to assess overall model performance^78,79. The multispectral model's OOB error rate was 7.75%, indicating a moderate risk of misclassification. The class-level F1 scores in the validation dataset ranged from 0.845 in TG to 0.984 in OP plant communities, indicating moderate variability in performance across communities.

The second evaluated RF model was based on VIs and DSM derived from the RGB aerial survey, achieving 98.89% accuracy with an overall F1 score of 0.987. Its OOB error rate was only 1.14%, reflecting very high reliability. Class-level F1 scores were consistently high (0.986–0.993), indicating robust classification for all communities.

The variable-importance plot in R (Figure 8A) demonstrated each VI's predictive abilities for the Boreal coastal wetland plant communities. The most influential multispectral VIs as predictors were mNDVI and SR, which indicate strong sensitivity to canopy greenness and biomass. Similarly, GRDI, NDVI, NDVIre and SRre were also highly ranked, reflecting the role of red-edge and normalized difference indices in distinguishing vegetation types. In contrast, the RGB-based index and DSM variable-importance plot show that DSM-derived structural height model as a predictor is a key driver of classification accuracy (Figure 8B). Other important predictors include TGI, grayscale, and VARI, which leverage RGB reflectance variability. Thus, these results indicate that DSM and RGB-derived indices significantly improve classification reliability and reduce misclassification risk.

The entire geospatial dataset was saved as the GIS Project, a folder, including UAV survey and VI raster maps, the biophysical observation GeoPackage (ground truth), and plant community maps; all linked as a QGIS project (.qgz file). The project is suitable for offline use and enables interoperability across GIS platforms.

Figure 1: Study site. Photograph with a location pin on the global map, pointing at the island of Hiiumaa in the Western Estonian archipelago. The study site, Kõrgessaare rannaniit, is located in Viscosa village and is a nature conservation area. Habitat type: boreal Baltic coastal meadow (Natura 2000 code 1630). Please click here to view a larger version of this figure.

Figure 2: Conceptual workflow diagram. (1) UAV-based aerial survey, (2) ground-truth data collection and georeferencing, (3) vegetation index computation, and (4) supervised classification using random Forest in R. Actual maps from the protocol implementation illustrate this four-phase method. Please click here to view a larger version of this figure.

Figure 3: Detailed workflow diagram. The diagram lists the 4 phases (and key steps) of the protocol in different-coloured panels, along with datasets generated in each phase. Arrors symbolize data flow. Blue color distinguishes multispectral dataset from RGB in red. Please click here to view a larger version of this figure.

Figure 4: Photogrammetry software example. (A–C) A screenshot of an RGB image mosaicking project: (A) Processing options for DSM and orthomosaics. (B) Processing options for red, green, and blue indices. (C) View of the processing progress, including aerial images and camera angles. Please click here to view a larger version of this figure.

Figure 5. Clipping map raster files in GIS software. Screenshot of a QGIS workspace with visible layers and opened tabs in sequence demonstrating how to clip unnecessary data: raster-extraction-clip raster by Extent ("extent" in this case refers to a polygon shape drawn as desired and used to mask pixels falling outside its boundaries) Please click here to view a larger version of this figure.

Figure 6. GIS software Raster Calculator screenshot. Modified soil-adjusted vegetation index (MSAVI) formula input as an example. Please click here to view a larger version of this figure.

Figure 7. Statistical computing environment for R version screenshots demonstrating final outputs plotted in the R console (RStudio). (A) multispectral and (B) RGB-based plant community maps. Please click here to view a larger version of this figure.

Figure 8. Statistical computing environment for R screenshots demonstrating VI performance rankings. (A, B) Views of an R console: (A) Screenshot when using multispectral VIs as an input dataset, and (B) Screenshot when using an RGB+DSM dataset. Both (A and B) views include variable-importance plots named "final_rf" (script-generated object name). Please click here to view a larger version of this figure.

Supplementary File 1: Vegetation index equations. List of multispectral and RGB-based vegetation index (VI) equations used in this protocol, including formulas for RGB band normalization prior to RGB VI calculation. Please click here to download this file.

Supplementary File 2: RF classification and validation script.R. Please click here to download this file.

Discussion

The critical steps in the protocol are in phase 1, the UAV survey, and in phase 2, fieldwork, where biophysical observations are collected, and the ground truth dataset is compiled. Both are weather-dependent and must be executed in close proximity. And both must yield geospatial datasets with high coordinate precision (1–3 cm precision was achieved for the case study). The success of vegetation index calculations and the machine learning process directly depends on the data quality of the UAV Survey and ground truth. Crucial steps to achieve UAV-survey map accuracy are described in step 1.4, Post-flight procedure. If post-processing/image geotagging uses RINEX data that is not accessible, an alternative is to use a live RTK base station, which provides real-time GNSS correction signals to the drone during flight (see an example in the video article Movchan et al.³⁷ with an example for setup instructions).

Before data collection, the study site should be visited to gain experiential knowledge of the landscape, identify the plant communities of interest, and determine which biophysical observations can be feasibly collected on-site to describe plant community composition. Because each study location exhibits distinct landscape characteristics and flora shaped by its climate zone, identifying plant species is essential for accurate land cover mapping, as species-specific biochemical, structural, and physiological traits produce unique spectral reflectance signatures. For each sampling quadrat, coordinates should be recorded at all four corners using GNSS/RTK equipment and land‑surveying software, with the quadrat identification code entered and saved. All additional biophysical observations or samples collected within the quadrat must be linked to these coordinates so that every measurement is anchored to the geospatial dataset and can be compiled in a uniform tabulated format for plant community classification and subsequent statistical analyses or ecological modeling. The sampled quadrats are labeled as training points, with community names/codes attached (for example, in this training set, they were OP, LS, US, and TG). Random Forest (RF) then learns the pixel patterns associated with each label. The final result of RF classification is a map displaying the distribution of plant community labels across the study site, based on quadrat training samples provided during fieldwork.

Phases 3 (Vegetation index calculations) and 4 (Supervised classification in R using Random Forest) are interdependent and highly modifiable. For example, Supplementary File 1 lists 19 mathematical formulas for calculating indices from multispectral bands and 23 formulas for recalculating RGB aerial survey data. Nonetheless, as VIs are formulated by researchers, the equations are open source and published in peer-reviewed articles and spectral index databases⁸⁰. A greater number (than provided) of formulas are available for testing in the machine learning process to classify plant community maps. Changes to the datasets will affect the algorithm's performance. VIs can be added and omitted to test models. The availability of two final plant community maps, derived from multispectral and RGB-based datasets, provides an opportunity to compare model performances. The RGB model achieved 98.89%, but that does not necessarily mean the sensor is superior to the multispectral sensor, which achieved a lower score. RGB-based data relies on color and height, which can be efficient for classifying landcover maps of structurally diverse habitats^9,33. In contrast, multispectral imagery provides more information about plant physiology, not just appearance, and may be superior at distinguishing vegetation types with similar colors and shadow resistance (e.g., NIR light, which plants strongly reflect, can be captured even when they are partly shaded⁸¹). Using only RGB sensors is limited by their narrow spectral range (red, green, blue), as subtle biochemical and physiological differences (e.g., chlorophyll content and water stress) may not be detectable without NIR, red‑edge, or hyperspectral bands.

Main limitations of the method may include the costly UAV system and the photogrammetry demonstrated. However, the method is designed so that, to replicate this protocol, a multispectral sensor and an RGB camera can be used interchangeably. To implement this protocol, the eBee X fixed‑wing UAV system equipped with a multispectral and RGB sensor (Sequoia) and the eMotion flight‑management software was used; however, other UAV systems can reproduce the protocol provided if they include a multispectral sensor and/or an RGB camera, a mission‑planning function in the flight‑management software, a 90‑degree nadir camera orientation, and either a user‑accessible flight log for post‑flight geotagging or an RTK capability to record precise image coordinates during flight. Furthermore, non-pilots may subcontract UAV surveys if the contracted pilot meets all legal and operational requirements, a common practice in research, conservation, and environmental monitoring; however, the research team remains responsible for fundamental mission design, including defining the study site and determining survey timing. Therefore, even for non-UAV pilots, it is recommended to familiarize them with the first phase of the protocol. In addition, the demonstrated photogrammetry software can be used to successfully implement the protocol, or it can be licensed for a short period to reduce costs.

A notable advantage of this protocol is its compatibility with both multispectral and standard RGB UAV cameras. It can be adopted by research teams and conservation practitioners with limited budgets and is scalable across regions and institutions, making it suitable for long-term monitoring. This hardware independence ensures the workflow remains future-proof and hardware-agnostic. Because both RGB and multispectral survey data can be used, the workflow supports comparisons across multiple sites, long-term time series—even when equipment changes—and the integration of legacy (historical) RGB datasets.

Future improvements to this protocol could incorporate texture metrics derived from high-resolution imagery to better capture structural differences in vegetation. In this context, exploring ensembles of RF models that combine RGB and multispectral imagery may improve classification accuracy. Also, future recommendations for upscaling UAV-derived plant community maps to satellite imagery and coupling this approach with cloud-based platforms would enable large-scale ecosystem modeling⁸². While UAV data remains essential for capturing the centimeter-scale structural and spectral detail required for accurate community-level mapping, satellite imagery could serve as an additional predictor layer or as broader contextual input. Incorporating satellite‑derived variables into the RF model may enhance its ability to generalize across larger spatial extents and support multi‑scale habitat assessments. This hybrid approach would allow UAV‑based community maps to be nested into a broader landscape perspective, enabling the method to detect regional patterns, prioritize UAV survey locations, and strengthen long‑term monitoring frameworks. Exploring how satellite data can be harmonized with UAV inputs within the RF pipeline represents a research opportunity to expand the scalability and operational utility of the protocol.

Integrating UAV imagery, field surveys, vegetation indices, and AI-assisted classification into a single, reproducible workflow is essential, as it provides the spatial, spectral, and ecological detail required for accurate mapping of low-growing vegetation. UAV-based remote sensing has gained increasing attention in the scientific community and has rapidly evolved into a widely used tool for diverse applications, particularly in coastal environment monitoring⁸³. This method provides a practical guide to a UAV-based, machine-learning-assisted workflow for education and research, intended to support ecological restoration and monitoring natural or semi-natural grasslands in coastal ecosystems.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
eBee X	senseFly/AgEagle	eBee X, https://www.sensefly.com/drones/ebee-x/	UAV system eBee X is a professional fixed-wing mapping drone. An alternative mapping drone system can be used
eMotion	senseFly/AgEagle	https://www.sensefly.com/drone-software/emotion/	Used to plan drone flight missions and manage aerial data. eMotion Software is a complimentary part of the eBeeX drone system
External hard drive	Any supplier	NA	Size depends on the study area size project size (250 GB - 1TB)
Fieldwork laptop	Any supplier	NA	a capable, portable machine suitable for field operations and running drone flightlifght management so (Based on eMotion specs:  64-bit Windows 10/11; Intel Core i5/i7 (12th Gen+) or AMD Ryzen 5/7 (>3.5 GHz) - High clock speeds are more important than core count; 16GB Minimum (32GB+ Highly Recommended); Mid-range NVIDIA GTX/RTX (e.g., RTX 3060/4060); NVMe SSD (Required).
GNSS/RTK system Trimble	Trimble Inc.	Model  R10	It has a receiver with land surveying software and spare batteries. Alternative GNSS/RTK systems can also be used.
Grass shears	Any supplier	NA	biomass sampling (optional)
Pix4Dmapper	Pix4D	https://www.pix4d.com/product/pix4dmapper-photogrammetry-software/	Photogrammetry software: Pix4Dmapper is available to license on a monthly or annual basis. An alternative photogrammetry software can be used to implement this protocol.
Plant identification guide	Any supplier	NA	Used to conduct biophysical observations (online, book or print-out local plant ID guides)
QGIS	QGIS	3.44.9 Solothurn	A free and open-source GIS software, avaliable at https://qgis.org/, alternative GIS software can be used to impliment this protocol
quadrat frame of 1 m x 1 m	Any supplier	NA	To be used in phase 2 for field work/sampling, i.e. to conduct biophysical observations, can be self-made
R	R-Software Services	version 4.2.0, https://cran.r-project.org	An open-source a predictito download modelling environment widely used for statistical computing and graphics, data analysis, and predictive modeling.
RStudio	R-Software Services	version 2024.4.0.735, https//posit.co/download/rstudio/
Rtools	R-Software Services	https://cran.r-project.org/bin/windows/Rtools/	As some packages may require compilation from source. Key pacages used: terra 1.5 21, sp 1.5 0, sf 1.0 15, rgdal 1.6 6, raster 3.5 15, rsample 1.3 1, MLmetrics 1.1 1, and randomForest 4.7 1.1.
Scale (battery-operated, levelled, precision ±1 g)	Any supplier	NA	durable for fieldwork
Sequoia	Parrot SA	Parrot SEQUOIA+	It is a Multispectral/RGB sensor, to be used in phase 1. RGB and/or multispectral sensor (many professional drones include one or both sensors). To implement this protocol used Parrot SEQUOIA+ (company Parrot SA)
Soil moisture reader	Any supplier	NA
UAV forecast weather app	Any supplier	NA	To plan flights used https://www.uavforecast.com/
WET sensor kit	Delta T Devices	WET-UM-1.6	Soil moisture sensor (optional)
Workstation	Any supplier	NA	Optimal specification needed for smooth processing of the photogrammetry projects and machine learning in R (specs based on Pix4Dmapper software requirements: Windows 10/11, 64 bits; CPU quad-core or hexa-core Intel i9/Threadripper/Ryzen 9/; GeForce GTX GPU compatible with OpenGL 3.2 and 2 GB RAM; Hard disk: SSD; Large projects (over 500- 1000 images at 14 MP): 32 GB RAM, 60 GB SSD Free Space.