On-Farm UAV Imagery: Centimetre-Scale Intelligence

The Resolution Step

A Sentinel-2 multispectral pixel spans 10 metres on each side. Inside that 100 square metres sits a mixture of canopy, soil gaps, weed species, standing water, and shadow. The satellite aggregates all of it into one spectral signature. The instrument is extraordinary at landscape scale: NDVI patterns across hundreds of hectares, multi-seasonal change detection, free revisits every five days. The remote-sensing Sentinel-2 spoke covers that layer. What the satellite cannot do is resolve the individual plant.

The DJI Mavic 3 Multispectral delivers a ground sample distance (GSD) of approximately 3.0 centimetres at 30 metres above ground level (DJI product specifications, 2024). At that resolution, a crop row is not a spectral average; it is a sequence of discrete plant canopies. Leaf morphology separates species. Broadleaf weeds among cereal stems are classifiable by their leaf outline, not inferred from reflectance statistics. Stroppiana et al. (2018, Precision Agriculture) demonstrated weed species classification accuracy above 80% for seven weed types in maize at 4.9cm GSD using multispectral UAV imagery and random forest algorithms. The classification requires individual-plant resolution. Sentinel-2 cannot provide it. A drone at 30 metres AGL can.

Disease lesions operate at the same scale. Mahlein et al. (2012, Precision Agriculture) demonstrated that Septoria tritici blotch in wheat was detectable 7 to 14 days before visual symptoms appeared when monitored with fine-resolution multispectral imagery. A chlorotic patch visible from the field margin is already a late-stage event. The infection detectable at centimetre scale in the red-edge band is early-stage, when a fungicide application is still targeting a small, locatable area rather than a field-wide outbreak.

The aerial monitoring and regen verification spoke addresses satellite and drone-based monitoring at the broader verification scale. This spoke is scoped to the specific on-farm applications that centimetre resolution unlocks: plant-level weed detection, early canopy disease mapping, thermal stress localisation, and the stitching pipeline that turns raw imagery into actionable prescription layers.

Fixed-Wing vs Multirotor: The Coverage Trade

The platform choice is a coverage-versus-manoeuvrability trade with a direct farm-size implication. Multirotor drones hover, hold position, and fly tight grid geometries over irregular field shapes. The DJI Mavic 3 Multispectral carries a four-band multispectral sensor (green, red, red-edge, near-infrared) plus a 12-megapixel RGB camera. At standard grid-mission settings, 30 metres AGL, 70% side and forward overlap, a single battery covers approximately 15 to 20 hectares in a 25-minute flight. Retail price is $3,199 USD (DJI, 2024). For a farm of 50 to 150 hectares, three battery cycles in an afternoon cover the full field at 3cm GSD. The certification and launch requirements are those of a consumer product.

Fixed-wing platforms trade hover capability for endurance and coverage. The eBee X, manufactured by senseFly (a Parrot Group subsidiary), pairs an interchangeable sensor payload including the WX Multispectral (4-band) and S.O.D.A. 3D RGB options. At 150 metres AGL with the WX sensor, the eBee X delivers 3.7cm GSD across a maximum survey area of approximately 500 hectares per flight, using a catapult launch and belly-land recovery (senseFly, 2024). Complete system pricing with one sensor sits in the range of $16,000 to $22,000 USD depending on configuration. The Wingtra One GEN II (Wingtra AG, Switzerland) offers a vertical-takeoff-and-landing fixed-wing hybrid with a 42-megapixel Sony RGB and optional multispectral payload at approximately $17,000 to $22,000 USD (Wingtra, 2024). Both platforms require a larger operating area for launch and recovery but reduce the per-flight time significantly on large holdings.

For spraying and seeding operations, the platform category diverges entirely. Agricultural spraying drones such as the DJI Agras T50 and XAG P100 are purpose-built for liquid and granule payloads, not high-resolution imaging. The drone spraying and seeding spoke covers that platform family and its economics separately. The imaging and spraying workflows complement each other: the Mavic 3 Multispectral generates the prescription map; the Agras T50 executes it.

Three Applications the Resolution Unlocks

Weed detection and prescription mapping

At 3 to 5cm GSD, broadleaf weed species are classifiable by leaf shape, venation, and canopy geometry. A wild mustard plant among cereal stems has a different spectral signature and a different physical outline from the surrounding crop. The weed map output from a Mavic 3 Multispectral flight, processed through Pix4D or Agisoft Metashape, identifies weed location and density at row level. That map feeds targeted herbicide application: reduced total volume, precisely located passes. Drone-derived weed prescription maps have demonstrated herbicide volume reductions of 10 to 30% in documented trials across German and UK cereal systems (Zheng et al., 2019, Computers and Electronics in Agriculture). The connection to precision drone application is direct: the imaging flight generates the variable-rate prescription that the spraying drone or ground rig executes.

Canopy disease mapping

The red-edge band (700 to 730nm) is the spectral region most sensitive to early-stage chlorophyll degradation from fungal infection. The DJI Mavic 3 Multispectral captures this band alongside green, red, and near-infrared. An early-stage Septoria blotch patch in wheat, or grey leaf spot in maize, is distinguishable in the red-edge band 7 to 14 days before the yellowing is visible to a walking scout (Mahlein et al., 2012, Precision Agriculture). That two-week window is agronomically significant: a targeted fungicide pass on a localised early outbreak costs less and performs better than a field-wide prophylactic application three weeks later. Individual plant monitoring also covers stand establishment: plant counts from a grid flight identify missing plants within rows, supporting informed decisions on targeted replanting.

Thermal imaging for irrigation stress

Water-stressed plants close their stomata, reduce evapotranspiration, and elevate canopy temperature by 2 to 5 degrees Celsius relative to adequately irrigated neighbours (Idso et al., 1981, Agricultural Meteorology). This temperature signal is captured by thermal sensors on UAVs. The DJI Zenmuse H20T, a dual payload combining a 20-megapixel RGB and a 640 by 512 thermal imager, is compatible with the Matrice 350 RTK body and retails at approximately $8,999 USD (DJI, 2024). A thermal flight following an irrigation pass locates zones where penetration was incomplete. The Crop Water Stress Index (CWSI) derived from thermal imagery identifies these zones at sub-field resolution, allowing targeted re-application rather than blanket additional watering across the whole field.

The Stitching Pipeline: From Raw Frames to Field Map

Raw drone imagery is not a field map. It is hundreds of overlapping frames, each georeferenced and carrying attitude data from the onboard GPS and IMU. A 20-hectare DJI Mavic 3 Multispectral flight at 70% overlap generates approximately 400 to 500 frames. The photogrammetry software matches features across frames, solves the camera geometry for each image position and orientation, and reconstructs a spatially accurate orthomosaic: a single georeferenced raster with consistent scale and projection across the whole survey area. The output is a standard GeoTIFF, openable in any GIS platform, compatible with FarmOS, and exportable as a variable-rate prescription layer.

Three software platforms cover the regen operator stack. Pix4Dmapper (Pix4D SA, Switzerland) processes locally or via cloud, outputs orthomosaics, NDVI and multispectral index rasters, digital surface models (DSMs), and point clouds. The agricultural subscription runs at approximately $350 USD per year (Pix4D, 2024). DroneDeploy (acquired by Autodesk in 2023) offers a cloud-first pipeline with automated processing and in-browser analysis at $1,500 to $3,000 USD per year depending on tier (DroneDeploy, 2024). Agisoft Metashape Professional (Agisoft LLC) runs as a perpetual licence at approximately $180 USD per seat (Agisoft, 2024), processes entirely on the operator's hardware without any cloud upload step, and produces outputs equivalent to Pix4D in orthomosaic quality and multispectral raster accuracy.

Processing time for a 20-hectare flight runs 45 to 90 minutes on a current-generation laptop with dedicated GPU. A 500-hectare eBee X flight, generating 2,000 to 3,000 frames, takes 4 to 8 hours locally or 1 to 2 hours through cloud processing. The FarmOS platform accepts standard GeoTIFF imports natively, allowing an operator running FarmOS as their farm record system to integrate the week's drone output into their spatial record library, tied to date and management decision, without moving data outside their own infrastructure.

Regulatory Reality and Operator Economics

In the United States, any drone operation outside recreational hobby use requires a Remote Pilot Certificate under FAA Part 107 (Federal Aviation Administration, 2016 onwards). The certification is an online knowledge exam administered through PSI Exams. The exam fee is $175 USD as of 2024. No minimum flight hours are required. Farm operations below 400 feet AGL in uncontrolled airspace (Class G) need no additional waiver or airspace authorisation. Aircraft weighing over 250 grams must be registered with the FAA at $5 USD per drone. The DJI Mavic 3 Multispectral at 895 grams requires registration. The practical commitment for an operator who has not yet certified is one day of study and an afternoon exam.

In the European Union, drone operations are governed by Commission Implementing Regulation (EU) 2019/947. Drones under 250 grams fall in Open Category A1, requiring minimal certification and permitting flight near uninvolved persons. Farm survey multirotors in the 250 gram to 4 kilogram range fall in Open Category A2, requiring an A2 Certificate of Competency. Those above 4 kilograms fall in Open Category A3, requiring operations away from populated areas, a condition that most agricultural field survey work on open farmland naturally satisfies with standard separation from buildings and roads. Operator registration with the national aviation authority is required across all categories.

The operator economics arrive at a number that changes the argument. DJI Mavic 3 Multispectral at $3,199 USD, Pix4D agricultural subscription at $350 USD per year, FAA registration at $5 USD: total Year 1 cost of approximately $3,554 USD. At one flight per week across a 30-week growing season, covering 15 hectares per flight, that is 450 hectare-seasons of imagery. Amortised Year 1 cost: $7.90 per hectare. From Year 2, Pix4D subscription is the only recurring cost. At 450 hectare-seasons, the per-hectare figure drops below $1. An independent agronomist walk-through for scouting and disease assessment in the US market runs $5 to $15 per hectare per visit (2024 market rates). An aerial scouting service runs $8 to $25 per hectare per flight. The drone more than breaks even in its first season, and continues breaking even for years after the hardware is paid for.

The data sovereignty question enters at the software layer. DroneDeploy is a cloud platform: imagery uploaded to DroneDeploy processes on Autodesk infrastructure, subject to Autodesk's terms of service. Pix4D offers a local-processing option under the same subscription. Agisoft Metashape processes entirely on the operator's own hardware, with no upload step and no third-party terms of service touching the imagery. An operator who processes locally and stores GeoTIFF outputs in FarmOS retains complete control of their field intelligence within their own infrastructure. The Data Sovereignty spoke develops the full argument about what happens when field intelligence leaves the farm in a platform's terms of service. The stitching software is where that choice is made each time the drone lands.

A sensor in the field is a hypothesis about what the operator needs to know. At 3 centimetres per pixel, the hypothesis is specific enough to answer itself before the scout has finished the first row.