What "AI" Means in a Variable-Rate Workflow
A variable-rate application map — also called a VRA map, prescription map or Rx map — is a grid laid over the field in which every cell carries a rate. The spray drone or tractor reads the map and changes its output as it crosses the cells. The basics are covered in our guide to variable-rate application.
In DroneField the map itself is built by deterministic tools: the index is computed from your orthomosaic, the grid is laid out to your swath, and the zones are classified by fixed rules. Luna's role sits on top of that. It reads the statistics of the index map together with the context you give it — crop, growth stage, treatment, applicator — and suggests the decisions an agronomist would otherwise make by hand: which index to use, which zoning basis, and what doses per zone, each with an explanation.
The output is a draft plan, not an order. It opens in the same application-map panel you would use without AI, every value can be changed, and the PDF report carries an expert-approval block. AI suggestions in DroneField are decision support; the operator or agronomist decides.
Step 1: The Four-Question Wizard
Luna starts with a short wizard of four steps, because the same index map means different things for different jobs.
Crop
Which crop is in the field. Index values and their interpretation differ between, for example, wheat, corn and sunflower.
Sowing date or crop age
Growth stage decides whether the canopy is still open or has closed, which in turn decides whether NDVI is still informative or NDRE is the better choice, and whether soil is part of the picture.
Treatment
What you plan to apply — for example a nitrogen top-dressing, a foliar feed, a desiccant or a herbicide on weed patches. The treatment determines the zoning logic more than anything else.
Applicator
What will apply it — for example a spray drone or a ground sprayer. This frames the plan around the machine that will execute it, and the export formats that follow.
Step 2: Index Choice and Zoning Basis
With the context known, Luna picks the index for the job from what the flight provides. A multispectral survey from a Mavic 3 Multispectral or Phantom 4 Multispectral offers NDVI, GNDVI, NDRE, EVI and VARI; an RGB survey offers VARI, VDVI, NGRDI and ExG. Our guides to NDVI vs NDRE and RGB vegetation indices explain the reasoning behind those choices.
Then it chooses the zoning basis — the question each grid cell is asked:
Mean index
Each cell is classified by its average index value. This is the natural basis for fertiliser and foliar feeding, where the rate should follow differences in vigour or greenness across the field.
Living share
Each cell is classified by the share of live plants it contains, with the live/dead threshold computed automatically. This is the basis for desiccation and weed patches, where the real question is whether there is green, living plant material in the cell at all.
Choosing the wrong basis is one of the most common mistakes in manual VRA planning — a mean-index map for a desiccation job, for instance, smooths away the green patches that actually need the product. Getting this choice right is one of the most useful things the AI does.
Step 3: Dose Suggestions With an Explanation
For each zone, Luna proposes a dose and explains the agronomic reasoning behind it — why a weaker zone gets more or less product, why a dried-down zone can get zero, how the treatment and growth stage shaped the numbers. The explanation matters as much as the number: it lets you check the logic against what you know about the field, the product label and local practice.
Optionally, Luna can add a yield estimate and general nutrient advice. Treat both as orientation. They are derived from index statistics and the information in the wizard, not from soil samples, tissue tests or the field's history — so they are a starting point for a conversation with the farmer, not a substitute for those measurements.
Step 4: PDF Report and Expert Approval
Luna writes a PDF report that documents the analysis: the index used, the zones, the suggested doses and the explanation. The report contains an expert-approval block, so the agronomist or operator who reviewed the plan can sign it off before the treatment.
For contractors this report is also the hand-over document. With project sharing, the customer opens a link in the browser and can view the orthomosaic, index map, boundary and precision map, and download the prescription and the AI report without installing anything. Shares expire after 30 days. The AI analysis documentation and project sharing documentation describe both features in detail.
Step 5: Review and Edit the Application Map
Luna's plan opens in the application-map panel, where it becomes an ordinary, fully editable map — exactly as if you had built it yourself. Everything the AI suggested can be changed, and the map preview updates live while you edit.
The grid cell size (2–10 m, 8 m by default) should match the spray swath, and the grid can be rotated to the field's long edge so the drone flies along the cells. You can work with up to 10 zones, in l/ha or kg/ha, and see the total quantity per zone and for the whole field — useful for filling the tank and ordering product. A pixel filter that is independent of the colour scale, a minimum valid share rule, an edge buffer (overhanging cells so the drone covers the field edge) and a zone buffer let you fine-tune the result. The application map documentation covers each setting.
Step 6: Export to the Spray Drone or Terminal
Once approved, the map goes to the machine in its native format. For DJI Agras (T20, T25, T30, T40, T50, T100), DroneField writes a VRA precision map as an Rx GeoTIFF plus the field boundary as KML and Shapefile, to the drone or a USB stick. For XAG P40, P100 and V40, it exports spot-spraying zones or a VRA map as KML, as well as custom routes. ABZ Innovation L-series drones get the boundary with obstacle cut-outs and Rx zones; Hardi sprayers a VRA Shapefile; tractor terminals ISOXML V3 or V4. ESRI Shapefile, GeoTIFF, PNG, JPEG and KML are available for GIS work.
This is what turns AI crop zoning into a working tool rather than a pretty report: the plan lands on the spray drone software side in the same session, without redrawing anything. The full survey-to-spray sequence is described in our spray drone survey workflow.
What Is Sent to the AI — and What Never Leaves Your Computer
Processing in DroneField runs locally: the images, the orthomosaic and the index maps are computed and stored on your own Mac, Windows or Linux computer. Luna is an online service, so the analysis step does need a connection — but only index statistics are sent, together with the answers you gave in the wizard. The images themselves are never sent.
This matters for contractors working with customer data and for anyone on a slow rural connection: there is no multi-gigabyte upload, and no imagery of the customer's field sits on a third-party server. Luna is part of the subscription with a monthly quota; the rest of the workflow — processing, indices, manual application maps, exports — works without it, and offline.
Examples by Treatment
The same field can lead to very different plans depending on the job. These examples show how the zoning basis follows the treatment.
Nitrogen top-dressing
Zoning by mean index. Early in the season NDVI separates thin from well-established zones; after canopy closure NDRE usually resolves differences better. The rate logic — whether weaker zones get more nitrogen to catch up or less because their yield potential is limited — is exactly the kind of decision the explanation should make visible, and the agronomist should confirm.
Foliar feeding
Also mean index, typically with a smaller spread between zone rates. With an RGB-only drone, VARI or VDVI can serve while the canopy is open; in dense crops a multispectral index is more informative.
Desiccation
Zoning by living share. Cells where plants are still green get the desiccant; cells that have already dried down can get a reduced rate or zero. NGRDI or VARI from an RGB flight is often sufficient here, because the question is green versus dry, not subtle vigour.
Weed patches
Zoning by living share on bare soil, stubble or before crop emergence. ExG makes green plants stand out from soil, so cells with weeds get the herbicide and clean cells are skipped. Once weeds grow inside a green crop, no index can separate weed from crop — the map shows green cover, not species.
The Limits of AI Crop Zoning
AI-assisted planning is good at structure: matching the index and the zoning basis to the job, proposing a coherent set of zone rates, and documenting why. It cannot see what the index cannot see. A low-index zone may be nitrogen deficiency, drought, disease, compaction or simply a different soil — the map tells you where the difference is, not what causes it. Luna does not diagnose crop problems, does not identify weed species and does not know the product label, local regulations or the field's history unless you account for them.
That is why the workflow is built around review. Walk or scout the zones that matter, check the suggested doses against the label and your own experience, adjust the plan in the application-map panel, and sign the report. Used that way, AI variable-rate application planning saves time on the routine decisions and leaves the judgement where it belongs — with the operator and the agronomist. For the wider picture of how drone mapping supports these decisions, see drone mapping for precision agriculture.