01Motivation
The limits of wading through mud
Traditional tidal-flat surveys ask researchers to walk into the mud for hours and examine vast areas in fine detail. That is physically draining and time-consuming, and it is only possible during low tide, so opportunities are limited. Tides and weather add real safety risks. Accuracy suffers too: species are hard to identify, records get missed, and coordinates drift. Over large areas, results depend on how experienced the surveyor is, and the same flat surveyed by another person or at another time may not produce consistent results.
- Physical burden and labor-intensive time
- Accuracy limits
- Poor reproducibility
02The idea
Let the drone do the walking
Unmanned aerial vehicles use remote sensing to photograph wide areas at high resolution in a short time. Fitted with a multispectral camera and paired with vegetation indices such as NDVI, a drone can capture both the physiology and the structure of plants. The team set out to test it head-to-head against the field survey: UAV versus boots on the ground. Their hypothesis was that combining multispectral indices with texture and shape features, after correcting for tide and seasonal change, would produce a species-level map that matches the field survey at κ ≥ 0.6.
03Theory
Seeing beyond red, green and blue
A multispectral sensor records several narrow wavelength bands at once, not just visible light. Near-infrared shows plant vigor and biomass, since healthy leaves reflect it strongly, while the red-edge band reflects photosynthetic efficiency and stress. On an NDVI map, colors near red mark healthy vegetation and colors near blue mark bare ground, sand and mud. But two species can share similar NDVI values, so spectral data alone is not enough.
- NIR: vegetation vigor and biomass
- Red-edge: photosynthetic efficiency and stress
- NDVI: high for plants, low for mud and sand
04Theory
Texture and shape fill the gap
Even when band values look alike, the surface pattern can differ: an evenly green patch versus a mottled one. The team adds texture analysis using the gray-level co-occurrence matrix (GLCM), which turns patterns into numbers through four statistics: contrast, homogeneity, entropy and correlation. A second supplement is shape analysis, where object-based image analysis (OBIA) groups pixels into segments and measures each one's size and shape. Together, spectrum, texture and shape give the classifier three independent ways to tell species apart.
- GLCM: contrast, homogeneity, entropy, correlation
- OBIA: segment size and shape
05Study site
Three plots on Donggeomdo
The study area is Donggeomdo, a small island in Donggeom-ri, Gilsang-myeon, Ganghwa-gun, Incheon. The team chose three tidal-flat sites. Selection favored places where many traditional field studies had already been done, and deliberately included both simple and complex vegetation so the drone's performance could be compared across conditions. From June to August they studied how tidal-flat field surveys are usually done, narrowed candidate sites from earlier studies, identified the halophytes living there, and fixed the final three sites.
06Method
A 5 m grid drawn in QGIS
To make the field survey match the drone data point for point, the team used QGIS, an open-source geographic information system. They saved base imagery of the three sites from Google Earth Pro as KML files, imported them, outlined only the tidal flat as a polygon, and generated random points within it spaced 5 m apart both side to side and front to back. Each point became a place to survey on the ground. After tracking in the field, points that fell outside the study area or could not be reached were deleted. The drone used was a DJI Mavic 3M.
- Google Earth Pro → KML → QGIS
- Tidal flat selected as a polygon
- Random points at 5 m spacing
- Inaccessible points removed after tracking
07Fieldwork
What grows on the flats
Sites 1 and 2 were surveyed on August 10, 2025 and Site 3 on August 11. Common reed appeared at all three. Site 2 was the simplest, with four species: reed, Carex scabrifolia, Zoysia sinica and Limonium tetragonum. Site 1 added Phacelurus latifolius, Iris laevigata and Cnidium japonicum. Site 3 had a different community, with reed, Bolboschoenus planiculmis, Phacelurus latifolius, barnyard grass, cattail and two seepweeds, Suaeda japonica and Suaeda maritima. This range from simple to complex is exactly what the comparison needs.
08Next steps
From points to a map
Completed so far: choosing and refining the topic, setting the hypothesis, establishing the three sites, and the field survey and basic QGIS work. Still ahead: processing drone data, matching specific variable values to each species in QGIS, running a Random Forest classification, comparing it with the field survey, correcting for tide, and evaluating how efficient drone-based classification is. The team expects a map at κ ≥ 0.6, targeting 0.75. If that holds, it would show UAVs can replace or complement traditional field surveys in tidal-flat monitoring.
- Done: topic, hypothesis, sites, field survey, basic QGIS
- Planned: Random Forest classification, comparison, efficiency evaluation
