How the Agras T100 Turns Gusty Coastlines into Survey-Grade Imagery Without Losing a Propeller

META: Dr. Sarah Chen dissects antenna placement, nozzle physics, and RTK geometry that let DJI’s Agras T100 map wave-lashed shorelines at centimetre precision—even when the anemometer reads 14 m s⁻¹.

The South China Morning Post’s weather page called yesterday “a fresh breeze.” Out on the granite headland I was standing on, the ultrasonic anemometer clipped to my vest said 14 m s⁻¹—Beaufort 7, the edge of gale. Perfect, I thought, for the final validation flight of our year-long coastal-erosion project. I slid the Agras T100 onto its carbon-fibre launch rail, told the ground station to hold a 3 m s⁻¹ forward overlap, and watched the drone disappear into the salt haze. Forty-two minutes later it came home with 1 847 nadir frames, every one of them tagged RTK Fixed, 1 cm σ. The secret wasn’t brute wind-resistance; it was a chain of micro-decisions—antenna geometry, nozzle calibration, and a shutter release timed to the trough of every fourth wave—that turns a sprayer into a survey-grade coastal mapper.

Why coastal wind is a different beast

Inland turbulence is chaotic, but coastline turbulence is rhythmic. Waves shed clockwise vortices; sea cliffs bounce them back. The result is a 3 Hz pulsing load that can yaw a drone 2° in 200 ms—enough to smear a 1 cm ground-sample-distance image into a 3 cm blur. The T100’s spec sheet quotes 15 m s⁻¹ max wind resistance, yet spec-sheet numbers are measured in a clean wind tunnel. Real salt spray adds mass; every gram of brine that sticks to a prop tip shifts the centre of gravity outward by roughly 0.4 mm. At 2 200 rpm that translates into 11 N of extra centrifugal force, enough to trip the flight controller’s vibration floor and force a gimbal recalibration mid-mission—unless you anticipate it.

Antenna placement: the 2 cm rule nobody prints

Most operators mount the D-RTK 2 antenna dead-centre on the roof of the boat or the back of the pickup. On a moving vessel that is a 3 m phase centre swing, equal to a 9 cm east-west shift in every photo. Instead, I strap the antenna to a 1.5 m carbon mast clamped to the same bedrock benchmark I use for my total station. The mast puts the antenna 18 cm above the spray envelope and, crucially, 2 cm aft of the drone’s front props. That 2 cm is the difference between a clean L1 phase and one that is reflected off the carbon blades—multipath that can drop the Fix rate from 99 % to 87 % in seconds. During yesterday’s flight the onboard .ubx log shows 99.4 % Fix over 14.7 km, even while the gust sock 50 m away stood straight out.

Nozzle calibration as a photogrammetric tool

The T100 carries four ceramic XR110015 nozzles, factory-rated 1.5 L min⁻¹ at 3 bar. For spraying, that number governs droplet spectrum; for mapping, it governs forward overlap. Close the solenoids, remove the diaphragm pump, and the airframe becomes a 49 kg flying gimbal with precisely known mass distribution. We ran a static thrust bench: at 85 % throttle the residual vibration amplitude is 0.12 m s⁻² RMS, half that of the stock M300. That stability let us lock the P1 camera to 1/1 250 s, f/4, ISO 160—fast enough to freeze a wave crest yet slow enough to keep noise floor under 1.2 e⁻. The nozzle bodies stay on because they act as vortex generators, tripping boundary-layer airflow over the boom arms and cutting lateral drift by 6 % according to our wind-tunnel smoke visualisation. Spray drift physics, repurposed for shutter sharpness.

Swath width versus salt-spray cycle

A common error is to fly the coastline at right angles, chasing the shortest flight time. Waves don’t travel perpendicular to shore; they arrive at 12° off the longitudinal axis here. Fly perpendicular and every third image is framed at the instant a crest explodes—white foam raises scene luminance by 1.8 EV and the photogrammetric algorithm sees no texture. We rotate the swath 12° to parallel the wave train, then time the shutter to the trough. With a 56 m swath at 80 m AGL the P1’s 35 mm lens gives 1.2 cm GSD; the wave period is 7.4 s, so we set the trigger interval to 1.85 s—exactly a quarter-phase. The result is a sequence where every photo sees wet sand, not foam, increasing tie-point matches by 22 % in Metashape.

RTK Fix rate as a reliability thermometer

Survey pilots obsess over RMS values, yet RMS is a mean; it hides outliers. Fix rate is the thermometer that tells you when the model will fall apart. Below 95 % you start to see dome-shaped error surfaces in the dense cloud—classic sign of a slowly drifting integer ambiguity. Our coastal baseline is 8.3 km long, so we run the rover in FIX-and-HOLD with a 0.05 m validation gate. The T100’s redundant IMU fuses barometer, dual-frequency GNSS, and visual odometry; the logs show only three brief (0.8 s) Float events, all coinciding with 90° bank turns at the end of each transect. Conclusion: keep bank angles under 35° when you are within 200 m of the base station to protect the phase track.

Multispectral hindsight

We piggy-backed a five-band MicaSense onto the rear boom for a post-storm vegetation recovery pass. The trick is to trigger the bands sequentially, not simultaneously; the 90 ms stagger avoids USB bus saturation and lets each frame ride the same IMU attitude solution. Georeferencing error drops from 2.3 cm to 0.9 cm, good enough to measure dune-grass NDVI change within one tidal cycle. Coastal scientists usually accept 5 cm; we just gave them a five-fold improvement with a firmware toggle.

IPX6K and the sand-blasting reality

IPX6K sounds like marketing gloss until you watch 30 kt sand hit an uncoated circuit board. The T100’s aluminium-magnesium chassis has a 35 µm hard-anodised layer, twice the thickness of the old T20. After 180 flights over volcanic quartz beaches the boom arm paint is frosted but the board bay still reads <50 Ω between any two grounding points—no white-rust creep. The gimbal damping balls are another story; shore-side UV ages them to 65 Shore A in six months. Swap them every 100 flight hours or you will see a 0.4 pixel jitter that no amount of RTK can fix.

Workflow checklist you can copy-paste

Drive two stainless stakes into bedrock 1.2 m apart; mount the D-RTK 2 mast on the windward stake so your body blocks spray when you service the radio.
Power-cycle the base station twice; the second reboot forces a new ephemeris download and cuts the convergence time from 120 s to 38 s.
Load the .kml transects into DJI Terra with 80 % forward, 70 % side overlap, then export as .dat to the T100—Terra respects the 12° wave-angle rotation.
Calibrate nozzles off-drone; 1 % flow error equals 0.3 kg mass bias, enough to shift the hover throttle by 0.8 % and the vertical scale by 1.1 cm in the final model.
Launch into the wind, never downwind; the initial climb is where prop-wash meets oncoming gust—worst-case load case. Let the autopilot handle it, but keep your thumb on the mode switch.
Land on a silicone mat; salt crystals at 40 °C scratch the anodising like 600-grit paper.

What the logs told me this morning

Back in the lab I fed 4 721 images into Metashape, disabled CPs, and ran “high” on a 64-core Threadripper. The sparse cloud threw 512 million tie points; the report lists mean reprojection error 0.28 pixel, horizontal RMSE 0.7 cm, vertical 1.1 cm—both below the GSD. The secret sauce was that 99.4 % Fix rate; every Float second propagates a 3–5 cm ripple that multiplies across the bundle adjustment. Holding Fix kept the ripple from ever starting.

From spray tank to science platform

DJI built the T100 to put fungicide on bananas. Coastal scientists care nothing for bananas, yet the same boom stiffness that resists tank slosh also damps shutter vibration; the pump bay that holds 40 L of glyphosate now cradles five-band sensors; the RTK engine that once guided row spacing now measures cliff retreat to the centimetre. Re-purposing is not hacking—it is recognising that precision is fungible. If you can hold a nozzle within 2 cm at 15 m s⁻¹, you can hold a pixel within 1 cm at 80 m.

Wind-burned and salt-crusted, I packed the mast and antenna into their Pelican case. The coastline will look different after the next typhoon, but the benchmark stakes are still in bedrock, waiting for the next flight. When colleagues ask how we beat the gusts, I tell them the drone did not beat anything; we simply removed every hidden source of error until the wind had nothing left to attack.

If you want the raw logs or the Terra project file to replicate the geometry, drop me a line on WhatsApp—https://wa.me/85255379740—and I will share the dataset under Creative Commons.

Ready for your own Agras T100? Contact our team for expert consultation.