Weed detection from your own imagery
Send us the multispectral captures from your Mavic 3M and a field boundary. You get back a weed map, a spray prescription and a survey report — checked by a person before it leaves here.
What makes this different
Most weed detection available today is built on trained AI models, which learn what a weed looks like from example images. That's a sound approach, and it works best when your field looks like the ones it learned from.
We've taken a different route. Detection here is deterministic: measured spectral thresholds across the green, red, red-edge and near-infrared bands, with spatial cleanup on top. Every polygon in your output can be traced back to the specific index values that produced it, and the same imagery run again next year gives the same answer.
That's useful beyond the agronomy. Spray records for herbicides and other plant protection products have to be kept for at least three years, and we keep your orthomosaics and maps for the same period. A map that can be explained in year three backs up that record — for an assurance audit, or when you're checking whether a patch has spread.
And a person looks at every job before it goes out. We see that as part of the service, not a stopgap.
The process
Mavic 3M with RTK, at 30 m, 80/70 overlap, nadir. Cloud is fine. The full spec is below — it's short, and it's a gate, not a suggestion.
Send the captures, your field boundary, and the product or tank mix and rate you plan to apply. We check the upload before you're charged anything. If it won't process, we tell you why and what to change — and there's no charge.
Orthomosaic, index maps, classified weed raster, patch statistics and a prescription you can load straight into the sprayer terminal.
Flight specification
There's no weather window here, no calibration panel and no time-of-day rule. Detection looks at the relative difference between weed and crop within your flight, the Mavic 3M's sunlight sensor corrects for changing light frame by frame, and thresholds are set for each field. What the spec protects is the three things processing can't fix afterwards: geometry, a complete band set, and RTK position.
| Parameter | Requirement |
|---|---|
| Sensor | DJI Mavic 3M |
| Bands | G · R · RE · NIR |
| Positioning | RTK fixed ≥ 95% |
| Sunlight sensor | Fitted, unobstructed |
| Altitude | 25–40 m AGL |
| Forward overlap | ≥ 80% |
| Side overlap | ≥ 70% |
| Camera angle | Nadir (−90°) |
| Elevation optimisation | Optional |
| Field boundary | .shp + .dbf + .shx |
We'd rather reject an upload than process it badly and explain afterwards. A rejection comes with the reason and the fix.
| AGL | MS GSD | Accepted |
|---|---|---|
| 25 m | 1.15 cm | Yes — finest |
| 30 m | 1.38 cm | Yes — recommended |
| 40 m | 1.84 cm | Yes |
| 50 m | 2.30 cm | By arrangement |
| 60 m | 2.76 cm | No |
Everything is processed and delivered at native resolution. That's the point: at 30 m the Mavic 3M resolves individual rye grass tillers inside a wheat canopy, and averaging down would throw away exactly the small, early patches you most want to catch.
RTK is required because a prescription has to be right in absolute position, not just relative. Your sprayer places each nozzle by RTK; a map on standard GNSS is out by metres, which puts a 25 cm spray buffer on clean crop and misses the weed.
Deliverables
Evidence
The field was flown, the prescription went into the sprayer the next day, and the as-applied data came back from John Deere Operations Center: the sprayer treated 17% of the field.
Measured, not modelled. That 83% comes from what the sprayer actually did, not from the geometry of our map. Any figure we haven't got as-applied data for is labelled as modelled — and it's a fair question to ask of any savings figure you're quoted.
Prescription maps are accurate at the point of survey. Weeds that emerge after the flight aren't on the map. On that field a fresh flush came up between the survey and the glyphosate taking effect — that's UK weather, not a detection miss, but it's the reason to fly as close to your spray date as you can.
One more thing worth saying plainly: thresholds are set per field and per growth stage. A new field means a calibration step. It takes a fraction of the time the original development did, and it's what the person in the loop does on every job.
Pricing
Processing fields for your own customers? Volume rates and unbranded delivery are agreed directly.
Ask about contractor rates →You're invoiced once your upload passes our checks, and can pay by card or bank transfer. For a first job the maps are sent as soon as payment comes through — card payment means that can be minutes, not days. Regular customers can move to monthly invoicing.
Rejected uploads are free. If a job passes our intake check and then fails to process, that's our problem to fix or refund, not yours. What we don't refund is agronomic disagreement — the map tells you what was spectrally detectable from the air on the day, which is an input to your decision rather than the decision.
Questions
No drone?
MPA can fly the field with our own Mavic 3M and run it through the same process. Survey and processing are quoted together, depending on where you are and how much ground there is.
Ask about a survey →We're building a network of pilots who can fly fields to this spec, so farms further afield can get mapped without the travel. If that sounds like you, we'd like to hear from you. You'll need:
Get started
The sensible first job is a single field you already know well, so you can check the map against what's actually out there. Email for an upload link and the job sheet.