C-DRONE GUIDE · 1 SEPTEMBER 2026
Drone Weed Mapping: Site-Specific Spraying and Variable-Rate Maps in Arable Crops
In most arable fields, weeds are not spread evenly : they sit in patches — thistle rings, blackgrass at the headland, dock outbreaks — separated by wide clean areas. The sprayer, meanwhile, treats everything. In a trial by the French sugar beet institute, only 14% of a field's area was actually sprayed after drone mapping of thistle patches, for efficacy identical to a blanket treatment. In other ARVALIS trials, the saving on herbicide volume reaches 80-99% on lightly infested perennial weeds — and drops to almost nothing as soon as the infestation is uniform. Here is what a drone flight can really detect, at which growth stage it must be flown, what an application map usable by a sprayer console actually contains, why the gains vary so widely from field to field, and what the service costs in 2026.
Published on 1 September 2026, reviewed on 1 September 2026 — regulations in force as of September 2026.
A weed is a green plant in the middle of a green plant
Weed mapping is not a variant of vigour mapping. A standard vegetation index — NDVI, NDRE — measures chlorophyll activity : it separates vegetation from bare soil very well, but it does not separate a weed from a maize plant, both being active vegetation. That is the fundamental difference with variable-rate nitrogen, which exploits precisely the average signal of a uniform canopy and is content with decimetre or even metre resolution.
Spotting a weed calls for an entirely different logic, based on three combined cues. The first is geometric : the crop is drilled in rows, the weed is not. An algorithm reconstructs each row axis from the orthophoto, then flags as suspect any green mass sitting in the inter-row. The second is morphological : leaf shape, patch compactness, texture — which means seeing the seedling, and therefore a resolution of the order of one centimetre. The third is spectral : some species separate from the crop on the red, red-edge and near-infrared channels, and that signal can help where geometry no longer suffices.
All three cues share one enemy : canopy closure. As long as soil is visible between rows, geometry works and seedling shapes are legible. Once the inter-row closes, the weed is mixed in with, or plainly hidden under, the crop, and no aerial imagery — drone, aircraft or satellite — will find it. That is also what separates this exercise from invasive plant mapping in natural habitats, where the target species often dominates the canopy and can be spotted by its own texture across a whole season.
The feasibility of the approach has long been documented. J. M. Peña, J. Torres-Sánchez, A. I. de Castro, M. Kelly and F. López-Granados published in 2013 in PLoS ONE a fully automatic object-based image analysis procedure applied to drone images of a maize field : flight at 30 m, 2 cm resolution, crop at the 4 to 6 leaf stage, with an overall classification accuracy of 86% and a coefficient of determination of 0.89 between estimated and observed weed densities (see the study on Google Scholar). The most interesting figure in that study, for a grower, is not the accuracy : it is that 23% of the field carried no weeds at all and 47% stayed below 5% weed cover. In other words, half the field did not need treating.
The flight window: a few days, not a few weeks
This is the constraint that shapes the whole service, and the one buyers underestimate most. The map is only worth anything if soil is still visible and the weed is already developed enough to be seen. Those two requirements pull against each other, and their intersection lasts a few days.
Research has quantified the trade-off. J. M. Peña and co-authors published in 2015 in Sensors a study systematically varying three parameters over a sunflower field : camera type, flight altitude (40, 60, 80 and 100 m) and date (44, 50 then 57 days after sowing, i.e. 4 leaves, 5-6 leaves then 7-8 leaves). The best detection accuracy, up to 91%, was obtained with colour-infrared imagery at 40 m on the second date, at the 5-6 true leaf stage (see the study on Google Scholar). One counter-intuitive detail : at low altitude, flying earlier gave better results, whereas in visible imagery at 60 m and above the third date won — simply because the weeds had grown and finally became visible at that resolution. Resolution and calendar partly compensate for each other ; but you cannot make up for flying too high by flying too late without losing the agronomic value of the map.
On top of this biological window sits an operational one : post-emergence weed control has its own dates, and the application must closely follow the flight. A map produced on 3 cm seedlings and applied three weeks later describes an infestation that has changed — fresh flushes have emerged, patches have spread. In practice, the reasonable delay between flight and application is a few days : hence the need to process the data within 24 to 72 hours and to book the sprayer slot before the drone even takes off.
One last factor: the weather. A 40 m flight with high overlap needs moderate wind and stable lighting ; two days of rain or gusts can be enough to miss the window. This is one of the rare cases in precision agriculture where the drone keeps a decisive edge over satellite imagery — take off on demand between two showers, with no dependence on an orbital pass or a gap in the clouds, a trade-off detailed in our comparison of drone versus satellite in precision agriculture.
Centimetre RGB or multispectral: two approaches, two flight economics
Two families of methods coexist, and the choice is made in the field, not in the contractor's catalogue.
The first relies on very high resolution RGB. You fly low with a good visible sensor to get a ground resolution of about 1 to 2 cm, then discriminate by out-of-row position, shape and texture. It is the cheapest approach in hardware — a standard inspection drone will do — and the finest for spotting isolated seedlings, but it demands heavier image processing and degrades quickly if rows are irregular or crop residues clutter the scene.
The second adds spectral bands — red-edge, near-infrared — to exploit reflectance differences between species. It is more robust where geometry falls short, and it produces the best published accuracies. It carries an area cost, though : in the 2015 Sensors study, each visible-camera image covered roughly 4.6 times more ground than a multispectral one — 0.28 ha against 0.06 ha at 40 m altitude. For a given area, a multispectral campaign therefore needs far more flight lines, hence more batteries and more time. Our guide to choosing between a multispectral and an RGB camera on an agricultural drone sets out what each sensor brings, and the one on ground sampling distance (GSD) explains how altitude, focal length and sensor size determine it.
Over large areas the airframe matters too : at 40 m with high overlap, a multirotor quickly hits a ceiling in hectares per day. A fixed-wing drone covers more, at the price of a launch area constraint and less flexibility on small fragmented plots. That trade-off belongs in the brief, based on the actual field layout rather than a theoretical total area.
The same flight, if properly set up, often serves twice : images acquired at the early stage also support emergence counting and plant density assessment, a decision taken in exactly the same window. Bundling both deliverables into one pass clearly changes the economics.
From map to console: format, cell size and software limits
A weed map is only operationally useful if the sprayer can read it. The expected deliverable is therefore not an image but an application map : a file of georeferenced zones carrying a rate — full rate on the patches, zero elsewhere, or several steps if the machine modulates.
Two formats cover most of the French fleet. The Shapefile (.shp and its companion files) is the most universal. ISOXML is the ISOBUS format : a TASKDATA folder holding a TaskData.xml file and a same-named .bin file carrying the rate grid. Each brand adds its own naming requirements : on John Deere the file must be called Rx on the USB stick ; on Trimble (GFX/TMX) the map goes into an AgData folder then Prescriptions, with prior conversion if the source is ISOXML. On ISOBUS, finally, the TC-GEO option must be enabled on both console and implement, failing which the map is ignored with no explicit error message. These details are trivial once known and cost half a day when they are not : the console make and model belong in the mission brief, not in the delivery.
Cell size is set by the sprayer's shut-off capability, never by image resolution. A 20 cm grid on a boom that shuts off in 4 m sections is pointless : the console will round it. Two configurations dominate : section control, originally designed to avoid headland overlaps and repurposed here to confine the treatment, with a useful cell size of 2 to 6 m ; and nozzle-by-nozzle shut-off, which brings it down to about 50 cm. A serious contractor also applies a safety margin around each detected patch : the zone is buffered by a few metres to absorb positioning errors and undetected seedlings, which increases the treated area but protects weed control efficacy.
That margin is no theoretical precaution. In the ARVALIS Digifermes trials, absolute positioning accuracy proved the sticking point, with discrepancies reaching ± 70 cm in some cases — enough to shift a patch treatment off target. It is also why the flight's georeferencing deserves discussion : our guide to the RTK or PPK drone explains when centimetre accuracy is justified and when it is a luxury.
One last limit, rarely documented by manufacturers : consoles cope badly with over-complex maps. The same ARVALIS trials put the threshold around 100 polygons and 5,000 nodes ; beyond that, application delays of 1 to 4 m appear, difficulties confirmed across several manufacturers. A map produced by fine-grained detection spontaneously yields thousands of micro-polygons : simplifying it — merging nearby patches, smoothing outlines, capping the object count — is part of producing the deliverable, and is what separates a usable map from a pretty GIS file.
The gains: why they range from 14% of area treated to no saving at all
Published figures on site-specific weed control are spectacular and widely scattered. The scatter is not a flaw in the studies : it describes agronomic reality exactly.
At the top of the range, the French sugar beet institute ran a 2021 trial of site-specific thistle control in sugar beet from multispectral drone images acquired at the 8-10 leaf stage : detected patches were automatically buffered for application safety, then exported as a variable-rate map. Result : only 14% of the field was sprayed, saving €34/ha on a €40/ha herbicide, for a drone pass and data processing cost of about €15/ha — with efficacy identical to the blanket treatment, the only two shortfalls observed coming from sprayer application faults rather than detection.
ARVALIS, which has been testing targeted spraying on perennial weeds since 2018 at its Boigneville and Saint-Hilaire-en-Woëvre research farms with three different sensors (multispectral and RGB carried by drone, RGB and hyperspectral carried by tractor), reaches converging orders of magnitude : 80 to 99% reduction in herbicide volume on lightly thistle-infested fields, depending on patch distribution and boom section width, and up to 88% of product saved in a 4.5 ha trial.
The review by R. Gerhards, D. Andújar Sanchez, P. Hamouz, G. G. Peteinatos, S. Christensen and C. Fernandez-Quintanilla, published in 2022 in Weed Research, takes as a consolidated order of magnitude at least 50% herbicide saving from patch spraying across various crops, with no extra weed control cost in following seasons — an important point, since the recurring objection is that the seedbank left in place takes its revenge the next year (see the study on Google Scholar).
That leaves the unfavourable case, which no sales brochure highlights : where infestation is uniform, the gain is nil. A field evenly infested with annuals emerging together will be treated in full, and the map will only have confirmed it. The economic reasoning is therefore simple, and it applies field by field : the expected saving equals the herbicide cost per hectare multiplied by the clean fraction of the area ; the service is only justified if that product exceeds the cost of the flight and data processing. On a €40/ha herbicide, at least half the field must be clean to break even on a €15-20/ha pass. On an expensive grass herbicide or a perennial clean-up at €80/ha, the bar drops sharply and the case appears from just a few isolated patches.
To those direct gains add two less visible but often decisive effects in 2026 : a lower herbicide treatment frequency index, which counts towards environmental certifications and supply-chain specifications, and traceability : an archived application map documents precisely what was treated, where and at what rate.
The limits to know before ordering
Four structural limits must be put to the buyer before signing, or disappointment is guaranteed.
A closing canopy. Wheat at advanced tillering, sugar beet that has closed the inter-row, maize at 8-10 leaves can no longer be usefully mapped. Detection collapses not because the algorithm is poor, but because the information no longer exists in the image. In those situations the only remaining route is real-time on-board detection on the sprayer, able to discriminate plant by plant as it passes.
Weeds under the canopy. Even within the early window, whatever grows under a crop leaf is invisible from directly above. The map therefore describes what is visible from the air, not the actual population : it is a map of patches, not an exhaustive inventory. Buffering the zones partly compensates for this bias, never fully.
Species discrimination. Separating « out-of-row vegetation » from the crop is a largely solved problem ; telling a grass weed from a broadleaf, let alone one species from another, remains difficult at centimetre resolution from a drone. Practical consequence : the map usually drives a single product at a single rate. Choosing the product remains an agronomic act, based on field scouting and on the known flora of the plot.
False positives. Stones, crop residues, volunteers, damp patches : ARVALIS trials record about 20% false detections on thistle, while identifying close to 99% of the thistles present. That setting is deliberate and sound : in weed control a false positive costs a few square metres of product, a false negative costs an untreated patch that will re-seed. A contractor promising detection with no false positives is describing a dangerous setting, not a performance.
Against these limits, the on-board solution is a legitimate alternative with different economics. ARVALIS costed a detection and regulation package enabling nozzle-by-nozzle shut-off at roughly €3,500 excl. VAT per linear metre of boom, with sensors every 3 m : up to −35% herbicide and €44/ha saved, a herbicide treatment frequency index falling from 1.48 to 0.96, detection of 70% in maize (27% false positives) and 77% on grassland dock (22% false positives) — the investment paying off from 25% of area treated. The drone does not compete head-on with that hardware : it costs a one-off service rather than a capital outlay, it applies to an existing sprayer fleet, and it additionally produces an archivable map. The choice comes down to annual area and number of passes, exactly as with buying a drone versus hiring a contractor.
Finally, a weed map is not confined to chemical control. It also guides targeted mechanical weeding — rotary hoe or inter-row cultivator concentrated on the dirty zones, passes avoided elsewhere — and serves as a diagnostic layer for planning rotations : a perennial patch recurring in the same spot three years running says far more about a soil or drainage problem than about a weed control failure. On grassland the same logic applies to dock outbreaks, alongside biomass and sward height estimation.
Buffer zones and neighbour charters: what the map does not exempt you from
A point of vocabulary first, because it regularly causes confusion : flying a drone to photograph a field is not aerial spraying. Article L. 253-8 of the French rural code bans, as a matter of principle, the aerial spraying of plant protection products, drones included, with narrow openings detailed in our guide to agricultural drone spraying and spreading. A mapping flight does not fall under that regime : it is an observation flight, subject to the ordinary rules of airspace and unmanned aircraft operation. The regulatory constraint bears on the application that follows, carried out on the ground by the sprayer.
And that application obeys the same untreated buffer zones as a blanket treatment. The decree and order of 27 December 2019, in force since 1 January 2020, impose safety distances measured from the property boundary of dwellings and places hosting vulnerable people : 20 m, non-reducible, for the most hazardous products (CMR 1 substances, endocrine disruptors), 10 m for orchards, vineyards, trees and shrubs, forestry, soft fruit, ornamental crops over 50 cm tall and hops, and 5 m for other crops — hence for most arable cropping. Under a départemental charter of commitments approved by the prefect and with approved drift-reduction equipment, those 10 m and 5 m distances can be cut to 5 m and 3 m respectively. The order of 25 January 2022 further set a non-reducible 10 m distance for CMR 2 classified products whose marketing authorisation sets no distance and for which no admissible update request had been registered with the French food safety agency by 1 October 2022.
On top of that come the « water body » buffer zones of the order of 4 May 2017, specific to each product : 5, 20, 50 or 100 m depending on the marketing authorisation. A 20 or 50 m buffer can be cut to 5 m by meeting both of two conditions : a permanent vegetated strip at least 5 m wide along the water body, and an approved means of reducing drift by at least 66%. The 100 m buffer is not reducible.
This framework translates directly into the deliverable, and that is where the contractor adds real value : buffer zones must be built into the application map as exclusion zones, rate set to zero, before export to the console. A thistle patch detected 2 m from a ditch must not open the nozzles. In practice this means having, on top of the detection, the exact field boundary, the reference hydrographic network and the neighbouring property limits — layers to assemble at briefing time, not the evening before the pass. Since applicable distances depend on the chosen product and on the départemental charter, the map is finalised after the product is chosen, which is one more reason to set the schedule with the agronomist when the flight is ordered.
This regulatory information reflects the texts in force at this guide's update date ; départemental charters and marketing authorisations change, and the distance actually enforceable is always the one on the product label and in your département's charter.
Method and 2026 prices
How a typical assignment runs, for an arable farm or a grower group :
- Briefing, one to three weeks ahead. Fields concerned and GIS boundaries, crop and sowing date, expected flora (patchy perennials or diffuse annuals?), sprayer console make and model, boom width and shut-off mode (sections or nozzle by nozzle), available regulatory layers. That meeting determines the flight resolution and the map cell size.
- Triggered by scouting. The flight is booked but not dated : it goes when the field reaches the useful stage and the weather allows, generally within a window of a few days.
- Flight. Programmed low-altitude plan with high overlap, centimetre RGB or multispectral depending on the brief. Expect 20 to 50 ha per half-day with a multirotor at this resolution, more with a fixed wing.
- Processing, 24 to 72 hours. Orthophoto, row detection, classification of weedy zones, safety buffering, polygon simplification, integration of exclusion zones, export in console format.
- Delivery. Application map (Shapefile and/or ISOXML named to the brand's convention), control orthophoto, table of areas per field with the treated fraction and the estimated product saving.
Orders of magnitude observed in France in 2026 (excl. VAT) :
- Weed mapping + application map : €12 to €25 per hectare depending on the resolution required, the sensor and how fragmented the fields are, tapering beyond a hundred grouped hectares. The rate is higher than for vigour mapping for a simple reason : at 40 m you fly three to four times more lines than at 120 m.
- Minimum call-out fee (travel, flight, processing, console export) : €350 to €600 — hence the value of grouping fields with neighbours, through the machinery ring or the cooperative, all the more so as the stage window is often shared across a whole district.
- Emergence counting added to the same flight : + €3 to €6/ha, since the data is already captured.
- Integration of regulatory exclusion zones (neighbour and water-body buffers, boundaries, ditches) into the map : usually included ; allow €100 to €250 of set-up in the first year if the field layers have to be rebuilt.
The payback calculation fits on one line and is done field by field : herbicide cost per hectare × the expected clean fraction of the area, compared with the price of the pass. On patchy perennials with an expensive product, the operation pays for itself at the first application ; on a uniformly weedy field it does not — and an honest contractor will tell you so before flying. The right starting point remains a diagnostic flight on one or two fields known for their thistle or dock rings, rather than a commitment across the whole rotation.
This service builds on our agricultural drone offer and complements the season's other mapping work, described in our overview of drones in precision agriculture. If your project concerns aerial application itself instead, see our drone spraying and seeding rates per hectare and their legal framework. For a quotation, request a quote stating areas, crop, target flora and above all your sprayer's console brand : that is what guarantees an export you can load straight away, with no last-minute conversion at the field gate.
Frequently asked questions
At which crop stage should the drone fly to map weeds?
As early as possible after emergence, while the soil is still visible between rows. That constraint drives the whole schedule : a weed is distinguished from the crop by its position outside the row, its shape and its texture ; once the crop closes the inter-row, all three cues disappear and detection becomes very unreliable. In maize the window matches the 4 to 6 leaf stage — the same as the post-emergence herbicide application. On sunflower, work by J. M. Peña and colleagues places the optimum around 50 days after sowing, at the 5-6 true leaf stage. In practice the mission is booked in advance and triggered by field observation, not by a date set three weeks earlier.
Do you need a specific sprayer to use a weed map?
Yes, and it is the point to check before ordering the flight. You need a console able to read a variable-rate map (Shapefile or ISOXML) and a sprayer able to shut off : either by boom section — the useful resolution is then the section width, 2 to 6 m depending on the boom — or nozzle by nozzle, which brings the cell size down to roughly 50 cm. On ISOBUS, the TC-GEO option must be enabled on both the console and the implement. Without variable-rate hardware the map still helps as a guide for manual passes or targeted mechanical weeding, but the automatic product saving does not materialise.
What herbicide saving can you realistically expect?
It depends entirely on the spatial pattern of the infestation, not on detection quality. On clearly patchy perennials — thistle, dock — under light infestation, ARVALIS measured volume reductions of 80 to 99%, and the sugar beet institute a treatment limited to 14% of the area, saving €34/ha on a €40/ha herbicide. The review by R. Gerhards and co-authors published in 2022 in Weed Research takes at least 50% saving as the order of magnitude for patch spraying, without extra weed control costs in following years. Conversely, on a uniformly weedy field or with annuals emerging evenly, the map will simply confirm that everything needs treating : the gain is then nil and the cost of the survey wasted. A diagnostic flight on a field known for its patches is the only honest way to settle the question.