Experimental design
The study area was located along a coastal section of the Red Sea, approximately 80 km north of Jeddah, Saudi Arabia (Fig. 1). Within the study area, mangroves consisting of Avicennia marina, along with patches of seagrass and coral reefs were present within 500 m of each other. The site has a tropical climate and receives less than 100 mm of rainfall annually. The tidal range in this area is generally < 30 cm. On March 18, 2021, when the dye tracing experiment took place, high (58 cm above lowest astronomical tide) and low (40 cm above lowest astronomical tide) tides occurred at 09:42 and 16:23, respectively. On that day, the wind direction was predominantly from the west-northwest. The wind speed, which was measured with a Gill WindSonic sensor (Gill Instruments Limited, Hampshire, UK) during image acquisition (09:55–15:50), increased gradually from 2 m/s at the time of dye release to 6 m/s from 13:30 onwards (see Supplementary Tables S1 and S2). The coastal area consisted of shallow water of between 0.5–1.5 m depth to the south of the mangroves, including substrates of sand, coral rubble with macroalgae, reef and sporadic seagrass patches. Coastal mangroves occupy a 200–300 m wide interface between the land and water. A natural channel of approximately 200 m width and with a depth of up to 16 m, leads into deeper water to the north (Fig. 1). The channel is bordered by sand and coral rubble on the eastern side and sand and some coral reef structure on the western side.
Prior to the field experiment, 22 ground control points were deployed evenly within the imaged study area in the water (11) and on land (11) to enable high relative geometric accuracy between the collected UAV image datasets. Six radiometric calibration panels in white, four shades of grey, and black were strategically deployed within the study area, so that they occurred within all UAV images collected during UAV hovering and flight surveys (Fig. 1). The reflectance of the six panels were measured with an ASD HandHeld-2 spectroradiometer (Malvern Panalytical, Malvern, United Kingdom) and used for radiometric calibration of the UAV image data [13]. Two temporary weather stations were set up (Fig. 1 and Supplementary Figure S1) to monitor air temperature and humidity at 2 m above ground and wind direction and speed at both 2 and 0.5 m height.
At two locations, one near the mangroves and one over a patch of seagrass, 300 mL of fluorescent dye (Rhodamine WT) was released at 09:50 and 10:07, respectively. Seawater samples for laboratory-based assessment of dye concentration were collected within the first hour of dye release. The time of each water sample collection was noted to facilitate identification of coincident UAV images. Different locations of water samples were selected to ensure a large range of dye concentration measurements. The exact sample locations were determined from the UAV photos, as the person collecting the samples was visible within the UAV photos. The Rhodamine WT dye concentration of the water samples was measured using a Cary Eclipse Fluorescence Spectrometer, which was calibrated with the same stock of Rhodamine WT dye used in the field study. The excitation wavelength of the fluorometer was set at 546 nm, and the fluorescence intensity at an emission wavelength of 580 nm (the emission wavelength at which calibration standards showed peak intensity) was recorded. Fluorescence intensity was translated into dye concentration measured in parts per billion (ppb) using a calibration curve made by Rhodamine WT standards (with deionized water) ranging from 5-200 ppb.
UAV image data collection and processing
The UAV imaging system used for mapping and tracing dye plume extent and concentration consisted of a gimbal-stabilized 20MP Hasselblad L1D-20c camera (Victor Hasselblad AB, Gothenburg, Sweden) replicated on two DJI MAVIC 2 Pro quadcopters (SZ DJI Technology Co., Ltd, Shenzhen, China). To enable initial continuous tracing of the dye plumes, at the time of dye concentration sampling, each of the two quadcopters were flown to an altitude of 400 m and placed in a hovering position above the location of dye release. This altitude resulted in a ground coverage of approximately 530 m x 350 m of the individual photos (Fig. 1), which were collected every 10 seconds. During hovering, the UAVs were moved horizontally up to 70 m in an east-west direction to minimize sun glint, while still ensuring full coverage of the dye plume and inclusion of the radiometric calibration panels in each photo. Four hovering flights were undertaken over the mangrove release site from 09:48 − 11:22, while three hovering flights were carried out for the seagrass release site from 10:06–11:14. Each UAV hovered over the release sites for approximately 20 min, with a break in between each flight of less than 5 min (including UAV descent, change of battery and ascent to 400 m altitude). The individual photos were geo-referenced to the UAV-based orthomosaic produced from the first flight survey (11:23 − 11:39) and based on the coordinates of the position of the GCPs within the orthomosaic. Together with the camera spectral database provided by Jiang et al. [14], an empirical line correction method [13] was used to convert the digital numbers of the photos to at-surface reflectance based on the field-derived spectrometer measurements of the radiometric calibration panels.
A total of nine UAV flight surveys were undertaken between 11:23 and 15:47 based on pre-defined flight lines using the Universal Ground Control Station (UgCS) Client application (SPH Engineering, SIA, Riga, Latvia) for autonomous data collection. The rationale behind changing the UAV data collection method from hovering to a flight survey approximately 1.5 hours after dye release was the unpredictable nature of the dye plume movement and dispersion over a larger area with time. Covering a larger area with a UAV flight survey ensured full coverage of both dye plumes. Both dye plumes were covered within 2–3 overlapping flight lines (within 3–5 minutes), ensuring little movement of the dye plumes during an individual flight survey. Each flight survey covered eight flight lines, with photos collected every 2 sec, at 300 m altitude and a speed of 8 m/s, with a distance between flight lines of 130 m, ensuring a forward overlap of 94% and a sidelap of 66%. The collected photos from each flight survey were processed in the Agisoft Metashape software (Agisoft LLC, St. Petersburg, Russia). The initial photo alignment was undertaken at high accuracy with the key and tie point limits set to 40,000 and 10,000, respectively. The coordinates of the positions of the GCPs were extracted from the orthomosaic produced from the first UAV flight survey and used for geo-referencing of all subsequent UAV flight surveys to ensure geometric alignment of all UAV datasets. Following the geo-referencing, a dense point cloud was produced at ultra-high density and aggressive filtering, and then used for producing a digital surface model, which the surface of the orthomosaic was based upon. Each orthomosaic had a pixel size of approximately 7.4 cm. Similar to the individual photos collected during hovering, an empirical line correction was performed to convert the orthomosaics to at-surface reflectance.
Mapping dye extent from UAV image data
Geographic object-based image analysis (GEOBIA) focuses on segmenting neighboring pixels in an image to form homogenous objects. As opposed to per-pixel analysis, GEOBIA allows object information to be used for classification, including statistical values of pixels forming an object (e.g. mean, standard deviation, quantiles, etc.), object area and shape, texture of objects, context information based on object location in relation to other objects, and hierarchical multi-scale approaches [15, 16]. A GEOBIA approach was applied to consistently map the extent of the dye plumes in the individual hovering photos and the orthomosaics. However, slight adjustments to some of the applied thresholds were required based on the changes in substrates underneath the dye plumes. Three sections in the images were identified with different substrates and/or water depths, including the area around the seagrass release point, and shallower area with sandy substrates around the mangrove release point, and the slightly deeper area to the east of the shallow area with a sandy substrate (Fig. 1). Initially, three additional indices were produced, including: Red / Blue; Red / Green; and (Red / Blue) x (Red/Green). The indices were based on visual inspection of the dye plume reflectance characteristics of the UAV data and the information provided by Clark et al. [17].
Based on the original UAV orthomosaic (Fig. 2a), a multi-threshold segmentation algorithm was first used to classify objects (representing the dye plumes) with a Red:Green band ratio > 1.5 and an object area larger than 3 m2 (Fig. 2b). This initial step identified those sections of the dye plumes with the highest concentration, and from which the initial objects were further grown into neighboring areas of lower dye concentration. Using a pixel-based object resizing algorithm, the initial dye plume objects were grown outwards pixel by pixel as long as they fulfilled the following criteria: Red / Blue > 0.85 and Red / Green > 1.02 (Fig. 2c). Applying the pixel-based object resizing algorithm again, the dye plume objects were further grown if they fulfilled the following criteria: Red / Green > 1 and (Red / Blue) x (Red/Green) > 2 (Fig. 2d). The second object-growing step required slight modifications of the (Red / Blue) x (Red/Green) threshold (equating to dye concentrations of > 1.1, 2 and 6 ppb) based on the dye plume locations in relation to substrate and water depth for the UAV-based orthomosaics. After this, all neighboring dye plume objects were merged, and unclassified objects that were fully enclosed by the dye plume objects were also classified as part of the dye plumes (Fig. 2e). Finally, the edges of the dye plumes were smoothed by filling object intrusions and shrinking objects extrusions based on a kernel size of 9 x 9 pixels (Fig. 2f). An outline of the object-based rule set applied for mapping the dye plume extent and concentration can be found in Supplementary Figure S2.
Mapping dye concentration from UAV imagery data
Dye concentration samples were collected while the two UAVs were hovering over the dye release sites. For each of the samples, a representative collection of 5 x 5 pixels was visually identified immediately in front of the person collecting the dye concentration samples. Hence, different UAV photos collected at the precise time of the dye concentration samples were used to eliminate the impact of the dye plume movement during sample collection. The radiometrically normalized UAV spectral band values of each collection of 5 x 5 pixels were averaged for the blue, green and red bands. In addition to the blue, green, and red UAV bands, the Red:Blue and Red:Green band ratios were produced as well as the interaction term of those two ratios, i.e. (Red / Blue) x (Red / Green) to relate to the derived dye concentration measurements (see Supplementary Table S3). Scatterplots and the associated coefficient of determination (R2) and root mean square error (RMSE) were produced to evaluate the relationships between field and UAV data. The best-fit equation between dye concentration and the interaction term was used to predict dye plume concentrations from the UAV image data. Thirty dye concentration samples (14 from the seagrass site and 16 from the mangrove site) were related to the collected hovering UAV photos.
Evaluation of the UAV-based maps
Manual delineation from visual inspection of the dye plumes was used for evaluation of the UAV-derived extent mapped from the UAV orthomosaics and individual hovering UAV photos. Using the manually delineated dye plume extent as reference data for the UAV-derived maps, commission and omission errors were calculated. In this case, commission error represented the percentage area that was incorrectly included in the mapping results (false positives), while the omission error characterized the percentage area that was incorrectly omitted in the mapping results (false negatives). The percentage of the committed areas was calculated in relation to the full areal extent of the mapped dye plumes, whereas the percentage of the omitted areas was assessed in relation to the full areal extent of the manually delineated dye plumes, i.e. the reference data.