Satellite, Drone and Variable-Rate Agronomy

Most UK farms already own some precision farming technology. The tractor has GPS, the combine logs yield, and there’s an app on the phone showing a satellite picture of every field. Often none of it changes a decision. The map gets looked at in March, the fertiliser spreader runs a flat rate anyway, and the yield data stays on a USB stick in the cab door.

This page is about closing that gap. It covers what each sensor can and can’t tell you, where the AI part actually does any work, how to cost a variable-rate job before you commit, and how to get a map from a laptop into a spreader without losing an afternoon to file formats. The numbers are UK numbers: hectares, kg N/ha, RB209 targets and current input prices. Prices move, so treat the costings as a way of working rather than a quote.

What each data source actually measures

A satellite image doesn’t “see” nitrogen deficiency or yield. It measures reflected light in a handful of wavelength bands, and software turns that into an index. If you know which source measures what, at what resolution and how often, you’ll stop expecting maps to answer questions they can’t.

The main point here is that NDVI saturates. Once a wheat canopy goes past a green area index (GAI) of about 3, NDVI barely changes, so a map from late April can look flat across a field that actually varies a lot. Indices that use the red-edge band (NDRE, REIP) keep separating canopies further into the season. Sentinel-2 has three red-edge bands. The platforms that use them for the GS31-GS37 nitrogen timings tend to be more useful than the ones that put a colourful NDVI layer on everything.

Where the AI actually earns its keep

You won’t be prompting a chatbot with any of this. The machine learning sits inside the platforms and does three jobs.

Cloud and shadow masking. A Sentinel-2 pass over Shropshire in February is often 70% cloud. Models trained on millions of images flag the cloudy and shadowed pixels so the GAI map doesn’t show a phantom thin patch where a cumulus was sitting. Platforms do this with very different skill. If your map has a crisp-edged low patch that moves between dates, it’s a cloud artefact, not a crop problem.

Turning reflectance into agronomy units. A raw index means nothing to a spreader. The platform’s model converts it into GAI, tonnes of dry matter or kg N already taken up by the crop, calibrated against field samples. Yara Atfarm, xarvio Field Manager (BASF), Hutchinsons Omnia, SOYL’s SOYLsense, Agrii’s platforms and the Climate FieldView and John Deere Operations Center ecosystems all do a version of this. They don’t all calibrate equally well on UK crops and varieties.

Zoning and weed detection. Clustering algorithms group pixels into management zones, typically 3-5 per field. On drone imagery, image-recognition models count plants or pick out blackgrass heads in June. Blackgrass patch mapping works best at about 2 cm resolution, flown when heads stand above the wheat canopy, and it’s the most dependable thing a drone does on an arable farm.

None of this replaces walking the field. What it changes is which bit of the field you walk to.

Variable-rate nitrogen: the biggest lever on arable

Nitrogen is the largest variable cost on most combinable crops, and the satellite-to-spreader route is cheap and mature, so this is where most farms should start. Our detailed walkthrough, Variable-Rate Nitrogen from Satellite Maps: A UK Field Guide, takes you step by step through choosing image dates, setting zone rates and checking the result at harvest. Here’s the short version, with a costed example.

There are two logics, and you need to know which one your platform uses:

1. Canopy management (early doses). Thin areas get more N so they reach a target GAI of about 2 at GS31. Thick areas get less, which cuts lodging risk and PGR spend. AHDB’s wheat and barley growth guides set out this approach.
2. Yield potential (later doses). Areas with consistently higher yields get more N, and areas that never yield, such as headlands, shaded margins or wet hollows, get less. This only makes sense with several years of yield maps behind it.

Using the wrong logic at the wrong timing is the most common way variable rate costs money.

Worked example: 30 ha of second wheat

The field is 30 ha of second wheat on a medium soil in Lincolnshire, with a total RB209-based plan of 200 kg N/ha. The second dose at GS30-31 is planned at 100 kg N/ha flat. A clear Sentinel-2 image from 4 March gives this GAI map:

The flat rate would have used 3,000 kg N, so the saving is 65 kg. With ammonium nitrate (34.5% N) at around £330/t, N costs about £0.96/kg, which makes the fertiliser saving roughly £62. That’s small, and you should expect it to be small. Variable rate moves N around the field much more than it reduces the total.

The return comes from yield and risk. If the thin zone yields 0.3 t/ha more because it built canopy in time, that’s 2.1 t at £180/t, or £378. If the thick zone doesn’t lodge, the benefit could be larger, but you can’t count it until you’ve seen it. Against that, a satellite N service at about £4/ha costs £120 for the field. On these numbers you’re ahead by about £320, or just over £10/ha, before any lodging benefit. That’s a typical, honest result for one timing on one field. Across 400 ha of cereals it starts to pay for a person’s time.

Check the outcome at harvest. Split the yield map by the same zone polygons. If the thin zone is still 1.5 t/ha behind, it has a constraint that N won’t fix (compaction, drainage, pH or blackgrass), and next year it should get less nitrogen, not more.

P, K and lime: the maps that quietly pay

The returns on variable-rate nitrogen are thin. Variable-rate lime is often where the obvious money is, because pH can vary a lot within a field and a mistake lasts five years.

A common setup is grid sampling at one sample per hectare (or per 2 ha on uniform ground) for pH, P, K and Mg. It costs roughly £12-£20/ha through SOYL, Hutchinsons, Agrii, Frontier or an independent agronomist, and it needs doing every four to five years.

Worked example: 18 ha arable field, target pH 6.5

The previous plan was a flat 2.5 t/ha of ground limestone, or 45 t in total. Grid sampling returned pH readings from 5.7 to 7.3:

Zone summary: Field "Top Bunny", sampled 14 Sept 2026
Zone   Area (ha)   Mean pH   Lime req (t/ha)   Lime (t)
A        6.1        7.1          0.0              0.0
B        7.2        6.2          2.5             18.0
C        4.7        5.8          4.5             21.2
Total   18.0                                     39.2

That’s about 6 t less lime, roughly £240 at £40/t spread. The bigger gain is that Zone C gets 4.5 t/ha instead of 2.5. At pH 5.8, barley and oilseed rape lose yield and phosphate availability drops. With the flat rate, Zone A would have been pushed towards pH 7.5 for nothing, which raises the risk of manganese deficiency, while Zone C stayed acidic.

Variable-rate P and K follow the same pattern. Indices at 0 or 1 get build-up rates, index 2 gets maintenance, and index 3 or above gets nothing that year. On a field with a history of FYM on one end, the potash saving alone can cover the sampling bill.

Variable seed rate

Seed rate is the second easiest win after N, and many drills already support it. The usual input is a soil EC map or a soil-brightness layer from bare-soil satellite imagery. Heavier, cloddier ground and headlands get more seed to make up for poorer establishment. Lighter, kinder ground gets less.

For winter wheat drilled in mid-October, a typical spread is 300 seeds/m² on good seedbed zones up to 400 on heavy patches. Variable rate doesn’t usually save much seed. The value is more even plant populations, which make the spring N decisions above cleaner. Seed rate maps are also one of the few prescriptions that rarely do harm, so they’re a good way to test the file workflow before you trust it with fertiliser.

Drones: scouting, mapping and spraying

A drone on a farm is mainly a scouting tool. A DJI Mavic 3 Enterprise or Mavic 3M, flown on an automated grid with DJI Pilot 2 or a third-party app such as DroneDeploy, gives you an orthomosaic map at a few centimetres per pixel. DJI quotes up to about 200 ha per flight day at mapping height, but plan on less in a UK wind.

Drones genuinely pay their way on arable and mixed farms in these jobs:

• Blackgrass and wild oat patch mapping in June, used to target spot spraying, fallow or a change of cropping on the worst areas next season.
• Establishment checks on oilseed rape after flea beetle or pigeon damage, when a decision to redrill has to be made in days.
• Storm and lodging assessment for insurance claims, with dated, georeferenced images.
• Drainage diagnosis from bare-soil imagery after heavy rain, which shows standing water and old tile runs.
• Livestock checks on large upland grazings, with a thermal camera on the Mavic 3T variant for finding ewes at lambing or stray cattle at dusk.

The rules aren’t optional. You need a CAA Operator ID and Flyer ID, and multispectral survey work under 2 kg usually fits the Open category if you stay below 120 m and away from uninvolved people. Spray drones such as the DJI Agras T50 are another matter. They’re well over 25 kg loaded, so they need a CAA Operational Authorisation in the Specific category. Applying a pesticide from the air also needs HSE permission and a product whose approval allows aerial use. Most UK farmers who want drone spraying hire a licensed contractor rather than doing it themselves, and that’s the sensible order of events.

Grassland, dairy, beef and sheep holdings

Livestock farms get the most out of satellites through grass measurement, not prescriptions. Platforms such as Pasture.io estimate grass cover (kg DM/ha) per paddock from Sentinel-2, and some fill cloudy gaps with Sentinel-1 radar. They send a weekly cover figure and growth rate to your phone. That lets you build a grass wedge without walking every paddock with a plate meter.

The catch is calibration. Satellite estimates can drift by 300-500 kg DM/ha, especially on swards with a lot of dead material or clover. A good routine is to plate-meter (a Jenquip or Grasshopper) four or five reference paddocks fortnightly and compare. If the satellite consistently reads 250 kg DM/ha high, mentally adjust and carry on. In a rotational grazing system you only need the ranking between paddocks to be right to pick the next one.

On a 150-cow spring-calving dairy unit, getting one more grazing round in before housing is worth more than any arable map in this guide. That’s a real target if the cover data lets you hold an average farm cover of about 2,000 kg DM/ha into October. Variable-rate fertiliser on grassland has a narrower use. The best candidate is lime on silage ground, where sampling often shows the field corners that never get slurry sitting at pH 5.5.

Getting the map into the machine

More precision schemes fail at this stage than anywhere else. The map works on the laptop, but the terminal won’t read it, the units are wrong, or the spreader applies the rate a field-width late.

The checks that matter:

1. Terminal and licence. On ISOBUS, the terminal needs Task Controller with geo-based rate control. The licence is often called TC-GEO and costs roughly £800-£2,000 if it wasn’t bought with the tractor. Trimble, Topcon, John Deere and Müller terminals each handle prescriptions slightly differently.
2. File format. Most platforms export shapefiles (.shp with its .dbf, .shx and .prj, which must travel together) and ISOXML (a TASKDATA folder). John Deere and Climate FieldView prefer to push files wirelessly to their own displays.
3. Product units. A rate of “125” means 125 kg N/ha to the agronomist and 125 kg product/ha to the spreader. For AN that’s a 65% under-application. Convert before export, not in the cab.
4. Look-ahead. Spinning-disc spreaders need a look-ahead of 1-3 seconds so the rate changes where the zone boundary is, not 15 m past it.

This is what a correctly built shapefile attribute table looks like for the wheat example above, exported as product rates of AN (34.5% N):

FID  ZONE     AREA_HA  N_KGHA  PRODUCT   RATE_KGHA
0    Thin       7.02     125   AN 34.5      362
1    Average   15.11     100   AN 34.5      290
2    Thick      7.87      70   AN 34.5      203

The RATE_KGHA column is the one the terminal applies. Before going to the field, load the file on the terminal while sitting in the yard and check it shows three zones in roughly the right shapes and the right rates. It takes five minutes, and most mistakes get caught there.

Costing a precision setup

Here’s a realistic first-year budget for a 300 ha arable farm that already has an ISOBUS tractor and a weigh-cell spreader:

Government support changes from round to round. The Farming Equipment and Technology Fund has funded GPS guidance, variable-rate controllers and yield-mapping kit in past rounds, so check the current round on GOV.UK before buying anything on its list, because you usually can’t claim for items bought before the grant offer. Environmental schemes have also included nutrient management planning actions that reward exactly the soil sampling and planning work set out here.

Against the costs, put realistic returns. That’s £8-£20/ha from N on cereals once you’ve learnt the system, a one-off gain from lime and P and K that can easily run to £20-£40/ha, and gains you can’t price in advance from catching problems early. Don’t budget for a 10% yield lift. Independent UK trials rarely show one.

Where precision schemes go wrong

• Trusting a single image. Compare at least two cloud-free dates before setting rates. A frost-scorched canopy in February reads thin for a week and then recovers.
• Zoning a field that doesn’t vary. If the GAI spread across a field is less than about 0.4, apply a flat rate. Making a variable map anyway only adds noise.
• Never checking. If you never compare zone yields at harvest, the scheme never gets better and you can’t tell whether it’s paying.
• Platform sprawl. Three free apps with three boundary sets for the same field means three slightly different areas and arguments at the spreader. Pick one master field boundary, ideally RTK-traced or taken from your RPA land parcel data, and use it everywhere.
• Ignoring headlands. Satellite pixels along the hedge mix crop and hedge readings. A good platform buffers them out, while a poor one marks every headland as thin and loads it with N.

A first season, month by month

September. Pick one platform and sort out your field boundaries. Grid-sample the fifth of the farm that’s due. Check the terminal licence.

October. Drill one or two fields with a variable seed rate from an EC or bare-soil map, keeping the setup simple, to test the file workflow.

February-March. Choose three cereal fields with visible variation. Run variable-rate N on the GS30-31 dose using the canopy-management logic. Leave a flat-rate strip, one spreader width wide, through each field.

April-May. Look at red-edge maps before the flag-leaf dose. Walk the thinnest and thickest zones with a spade. If you have a drone, fly any oilseed rape you’re worried about.

June. Map blackgrass patches by drone or by eye from the sprayer cab and mark them on the same boundary set.

Harvest. Calibrate the yield monitor on the first trailer, not the last. Clip yield data to the N zones and the flat-rate strips.

August. Compare the three fields zone by zone against their flat-rate strips. The number you need is the yield difference in the thin zone, measured against the strip running right through it. If it’s positive, next year you can roll the system out to the whole farm with real evidence. If it isn’t, the strip has probably found a soil or drainage problem that no satellite will fix, and that’s worth knowing too.