Seagrass meadow extent and meadow-scape was mapped using three alternative approaches at Midge Point, a coastal turbid water habitat, in the central section of the Great Barrier Reef, in September/October 2017. Approach 1 included mapping meadow boundaries and meadow-scape (including patches and scars) during low spring tides on foot within two sites (MP2 andMP3, each 5.5 hectare in area) using a handheld Garmin GPSMap 64s (accuracy ±1.5–3 m) on the 17 September 2017. Approach 2 was where the meadows were surveyed at low tide with observations from a helicopter (Robinson R44) on the 17 October 2017. The boundaries of the meadows were delineated by on-board observers and tracked by helicopter (at 25 ±5 m altitude) using the tracks setting on a handheld Garmin GPSMap 64s. Within these meadows, observational spot-check data was collected at an altitude of 1–2 m above the substrate, from three haphazard placements of a 0.25 m2 quadrat out the side of the helicopter at a number haphazardly scattered points (10 m2). Approach 3 used PlanetScope Dove imagery captured on 09 October 2017 coinciding as close as possible to the field-surveys, with 3.7 m x 3.7 m pixels (nadir viewing) acquired from the PlanetScope archive.
For Approach 1, fine-scale meadow-scape boundaries (patches or scars within 5.5 hectare area) were mapped for each site using the imported GPS track to create a polyline which was then smoothed using the B-spline algorithm and saved as a polygon. In Approach 2, meso-scale seagrass meadow boundaries were mapped from the GPS tracks and by on-screen interpolation based on geolocated spot-checks, field notes, and geotagged oblique aerial photographs acquired from the helicopter. For Approach 3, we created spatially explicit seagrass maps from PlanetScope Dove imagery, and conducted the classification using a machine-learning model (Random Forest) coupled with a Boot-strapping process (100 iterations). The final model predictions were then gathered into separate rasters, based on Bootstrap Probability thresholds of 60% and 100%. The final rasters were cleaned using a majority filter algorithm, to eliminate stray pixel predictions using a moving window between 3 and 9 pixels depending on the size of the imagery.
