How VAI Precision Farming turns drone orthomosaics into plant-health maps, targeted spray prescriptions and a paper trail insurers can actually use.
Vision AI · September 2026
The field is no longer a guess
A paddock looks uniform from the farm gate. From 80 metres up, it is a patchwork: a dry corner, a weed flush along the fence, a strip the planter skipped, a hail scar that does not show until the next week. The people who pay for that variability: growers, agronomists, lenders, and insurers still often work from a windshield survey, a handful of photos, and a date in a notebook.
VAI Precision Farming is a desktop application for the rest of that story. It takes drone photographs (or an already-stitched GeoTIFF), hangs a true-to-ground orthomosaic on a satellite basemap, and then lets you do the work that used to live in three different products: measure plant health, train a small on-field model to mark what you actually care about, estimate yield zones, compare one flight to the next, and export a map a sprayer or a claims file can use.
It is built as a local app. Rasters stay in project folders on the machine. Field identity: name, crop, boundary, season, which flight produced which map-- lives in a small local database. There is no farm data round-trip to a public cloud required to open yesterday’s mosaic and ask a better question of it.
What the application actually does
From a memory card to a map
Import Project is the stitch path. The operator points the app at a set of overlapping drone stills; NodeODM builds the orthomosaic, and optionally a DSM and DTM when the job is not run in fast-orthophoto mode. Progress is visible. The result is not a pretty picture in a gallery. It is a georeferenced GeoTIFF the rest of the tools can compute on.
Not every useful map starts in the app. Attach Orthophoto takes an existing GeoTIFF or a JPG/PNG plus world file and .prj and registers it as a flight on a field and season. Create Project is where the field is named, the crop is set, and the boundary is drawn on a satellite basemap so later numbers (hectares, class areas, yield totals) are tied to a real polygon, not a guess.
View Project is the working desk. The mosaic sits on a Mapbox satellite base that still holds together when you zoom in. A working-orthophoto picker chooses which dated flight is on the map when a field has more than one. Layers include the ortho, DSM, DTM, and the satellite base. That sounds simple. It is the difference between “we flew last Thursday” and “this is the Thursday map, on the ground, at this boundary.”
Plant Health: indices with an honest band list
Plant Health runs a vegetation-index engine on the current orthophoto. RGB-only indices such as VARI, GLI, EXG, and MPRI are available on a typical three-band (or RGB + alpha) drone mosaic. Indices that need near-infrared like NDVI, NDRE, and their cousins stay off unless the raster actually has those bands. A four-band NodeODM file that is only RGB plus an alpha mask is not treated as NIR. That is a quiet accuracy choice, and it matters if someone later reads the map as a nitrogen story.
The engine works blockwise, masks invalid pixels, stretches on robust percentiles, and writes a Float32 index raster plus a colour PNG. The legend speaks in classes people already use in the field: weak, moderate, healthy, with percentages and hectares. A satellite clip over the same boundary can publish a preview health map before a drone ever leaves the truck clearly badged as preview, not drone GSD so a season can start with a baseline instead of a blank.
The Magic Tool: labels in the paddock, not a black box in the cloud
This is the piece operators remember. Magic Tool is not a generic “AI button.” It is a supervised grid over the field you just mapped.
You draw or import a boundary. The app lays a lattice of cells: size, rotation, and offset under your control so the grid follows the rows instead of fighting them. You paint a few cells as Wanted and a few as Not wanted. Three and three is the floor; more is better. The source can be the orthophoto, a vegetation index, or both. After a handful of labels, a preview of the grid starts to make sense of the paddock. Hang tight, the map is being published is the line on screen while the model trains and every cell is classified.
What you get back is a zone map of Wanted cells, not a scatter of unexplained points. Brush and eraser let you fix the obvious misses. A buffer of one cell ring can fatten treat patches so a boom actually covers them. Dual opacity sliders keep the analysis and the prescription visible at the same time, and Reset to original restores the first prediction if a cleanup pass went too far.
Then the map becomes work:
Convert the dissolved polygons into a targeted operation or an inspection layer, not a point buffer pretending to be a spray zone.
Assign rates: treat litres per hectare on the unwanted class, a different rate (including zero) on the remainder. The panel shows treat hectares, tank fills, and percent saved versus spraying the whole field at the high rate.
Path planning: spot spraying lawn-mowers the treat patches and ferries off between them; variable-rate covers the field and writes a rate at each waypoint.
Export what the rest of the farm already speaks: GeoJSON, PNG, treat GeoTIFF, shapefile zip, KML, CSV waypoints, ISOXML, and DJI KMZ with a Placemark the controller can open.
The Magic Tool’s job is not to replace the agronomist. It is to turn twelve clicks in the paddock into a map the sprayer, the notebook, and the claims file can share.
That last sentence is the product idea. The model is trained on this field, this day, these labels. It does not pretend to know every weed in the district. It does pretend, correctly, that a georeferenced zone with an area in hectares is more useful than a memory of “the north end looked bad.”
Yield, terrain, and time
Yield Map uses the same georeferenced orthophoto and, when it exists, the Plant Health raster. Expected tonnes are field hectares (valid pixels inside the boundary) times a mean tonnes per hectare, optionally scaled so the field average matches a number the grower already believes. Zones can be seeded with ground samples. The point is not a satellite yield product. The point is a map that lives next to the health map and the Magic zones, in the same coordinate system.
Where a DTM was produced (bare-earth, Fast Orthophoto off), Irrigation maps slope, wetness-style accumulation, and ponding. It is terrain drainage, not a pump schedule, and the copy in the app says so. For insurance and agronomy alike, “water sat here” is often the missing sentence in a hail or drought file.
Compare flights tiles orthos or the same vegetation index across dates, graphs mean index over time, and can show Magic treat-percent then versus now. A difference map is the same index on two different flights, not two random layers on one mosaic. Side-by-side uses a shared colour stretch so April does not look greener than February only because the slider moved. That is how you talk about change without arguing about the legend.
Why this is useful when the conversation is insurance
Crop insurance, hail, drought, flood, and fire are not won on adjectives. They are won on where, when, how much, and compared to what. VAI Precision Farming was not written as a claims portal. It produces the kind of artefacts a loss adjuster, a grower, and an agronomist can put on the same table without translating three file formats by hand.
A dated, located record: not a phone album
Every mosaic is georeferenced. Every field can carry a drawn boundary and an area in hectares. Every flight can carry a captured-at timestamp and a source (stitched drone job, attached GeoTIFF, or satellite preview). Plant Health classes report hectares, not just colours. Magic sessions store the grid, the labels, the zone polygons, and the prescription. That is a chain: this field, this date, this map, this area of “not wanted,” this rate.
When a storm crosses a farm, the useful question is rarely “was there damage?” It is “how much of which block, relative to the week before?” A VARI or GLI map from a flight two days prior, and another after the event, is a before-and-after that still sits on the satellite base so anyone in the room can see the road, the dam, and the neighbour’s tree line. Compare flights and a same-index difference map are built for that sentence.
What a claims file can actually hold
Insurers and adjusters already accept photographs, farm maps, and agronomist letters. What they struggle with is unlocated JPEGs and screenshots with no scale. From this application you can put in the file:
The orthomosaic (or a PNG preview) with bounds: the field as flown, not as remembered.
A plant-health PNG plus the class key (weak / moderate / healthy, with hectares). That is a snapshot of canopy condition, with the honesty that RGB indices are not NDVI unless NIR was flown.
Magic zone GeoJSON or shapefile: treat versus remainder, with areas. Useful for hail streaks, weed blow-outs, and “this corner failed” in a language GIS tools already open.
A yield-zone map if the season had one, especially where ground samples were entered. Useful as context, not as a substitute for an insurer’s own yield model.
A DTM drainage map where ponding or runoff is part of the story.
A generated report (Word or PDF) that collects the maps instead of leaving them in a chat thread.
None of this replaces a site visit, a policy wording, or an adjuster’s measurement. It shortens the argument about which paddock and which week. It also helps the grower show mitigation: here is the spot-spray prescription we actually ran; here is the hectare we did not blanket with herbicide; here is the boundary we used.
Underwriting and season context, not only loss
The same stack helps before anything fails. A satellite health preview over a drawn boundary is a cheap baseline at the start of a season. A Magic session mid-season documents weed pressure with labelled examples, not a vibe. Yield zones, even as an internal estimate, give a lender or an insurer a picture of within-field risk that a single average yield does not. Multi-date index graphs show whether a block was already sliding before the named peril.
For parametric or index-adjacent products, the honest use is still documentary: the farm can show that a drone index moved when a weather index said it should, or that it did not — which is often the more important story. The application will not file the claim. It will keep the maps that make the claim specific.
What we do not to claim
RGB drone mosaics do not become NDVI because someone wants a familiar name. Irrigation from a DTM is not a water right. Yield from a vegetation index is an estimate scaled to a field average, not a weighbridge. Magic Tool is trained on the labels you give it on that flight; it is not a certified weed identifier for a whole region. Insurance markets differ; no output of this software is, by itself, a proof of loss. Used as dated, georeferenced, classed maps with exportable polygons, it is a much stronger annex than a folder of unlocated photos.
A day in the product
Morning: Create or open a field, confirm the boundary, attach last week’s GeoTIFF or stitch last night’s card. View Project, pick the working orthophoto if two flights exist. Plant Health, VARI, read the hectares in the weak class. If the question is weeds or a treatment map, Magic Tool: boundary, grid, a dozen labels, hang tight, cleanup, convert to layer, assign 80 L/ha on treat and 0 on remainder, export KMZ or ISOXML for the machine that will actually move.
After a storm: fly again or attach the new mosaic as a second flight on the same season. Compare. Same index, two dates. Difference map. Export the health PNG, the zone shapefile, and a short report. The conversation with the adjuster starts at a map, not at a disagreement about which corner of the pivot they meant.
See the field. Steer the work.
Precision farming software often stops at a pretty index. VAI Precision Farming is built to keep going: stitch or attach, inspect bands honestly, classify health with hectares, let you teach a grid what “unwanted” looks like on this paddock, turn that into a rate and a path, and keep every artefact on a named field with a date. Growers use that to spray less and scouting more. Agronomists use it to leave a map instead of a story. Insurers and the people who work with them can use the same files as evidence of where the crop was, what changed, and how much ground was involved.
The mosaic is the photograph. The Magic Tool is the argument you can export. The rest of the application is the filing cabinet that used to be a WhatsApp thread.