Yield forecasting

On this page 11

Count what is actually out there

Drones count plants, fruit or ears and turn the counts into a yield estimate for each part of the field.

CategoryDetect
Slugyield-forecasting
CadenceOnce at establishment, then at two or three points through grain fill or fruit set
PayloadHigh resolution RGB, multispectral, RTK
Needs a visit firstNo

The problem

Harvest planning, storage booking and forward selling all run on an estimate, and the estimate usually comes from a hand count in a few square metres extrapolated across a hundred hectares. A bad estimate books the wrong number of trailers and sells grain that is not there.

How it works

StepWhat happens
Fly lowA high resolution pass captures individual plants, ears or fruit at a density the model can count
CountObjects are counted per square metre and the counts are aggregated by zone rather than averaged flat
ModelCounts are combined with your variety, row spacing and historical yields to produce an estimate range
UpdateThe forecast is re-run on each later flight, so the range narrows as harvest approaches

What you get

  • Plant or fruit count per zone
  • Yield estimate with a range
  • Establishment gap map
  • Harvest sequencing suggestion

What the dashboard measures

  • Counted objects per square metre
  • Estimate range width
  • Gap area as a share of the field
  • Estimate against final weighbridge

The last one is the only honest measure of this capability, and it arrives once a year. Record it every year without exception: it is what makes next season's range credible.

Where it matters most

Winter wheat, maize, orchards, berries, cooperatives.

Build it

Two different products under one name

Be clear about which one you are building. They have different economics.

CountingEstimating
QuestionHow many plants, ears, fruit are hereHow many tonnes will come off
MethodObject detection over high resolution framesA model relating counts to weight
AccuracyGood, and verifiable immediatelyDepends on a season of calibration
Sells toAnyone deciding to redrill, replant or book pickersAnyone selling forward

Counting is buildable in a season. Estimating needs a calibration set of counts paired with weighbridge tickets, which is a second season. Ship counting, and be honest that estimates start wide.

Flight parameters

CropTargetAltitude
Cereal establishmentPlants per m²10 to 15 m, 0.3 to 0.5 cm/px
Cereal ear countingEars per m²5 to 10 m, or a ground rig
Maize establishmentPlants per row metre20 to 30 m, 0.7 cm/px
Orchard fruitFruit per tree8 to 15 m, oblique passes both sides of the row
BerriesFruit per plant, per flush5 to 10 m under tunnel headroom

Note the altitudes: ear counting is at the edge of what a multirotor can do safely over a standing crop, and orchard counting wants oblique views rather than nadir, because the fruit is on the sides of the canopy. Both push you toward a smaller aircraft flying transects rather than a survey grid.

Hardware

ItemWhyEU sourceUS sourceIndicative
High resolution RGB, global shutterCounting at 0.3 cm/px is unforgiving of motion blurSony ILX-LR1B&H€2,500 to €3,200
Compact alternative12 MP, tiny, good enough for maize and orchardsRaspberry Pi GS camera at BerryBaseAdafruit€55 plus lens
Small airframeLow, slow, close to the canopyHolybro X500 V2GetFPV€400 to €600
Precision rangefinderConstant AGL is what keeps the count per m² honestLightWare SF000/BRobotShop€200 to €300
Calibration framesQuadrats hand counted on the ground, the ground truth for everythingwire and paintsamenegligible
Gimbal for oblique workOrchard and vineyard counting needs a side viewGremsy or a fixed 45 degree mountGetFPV€0 to €900

In-house software

StageWhat we runLicenceReplaces
Detection or densityPoint-supervised counting (density map regression) rather than boxesown code, PyTorch BSDcommercial counting services
Tiled inferenceSAHI over full-resolution framesMIT
Overlap de-duplicationProject counts to ground coordinates, deduplicate in the overlapown code
AggregationCounts per zone using the zone layer from fertilisationown code
EstimationRegression from counts and indices to weighbridge tonnes, per crop and varietyscikit-learn, BSD
RangeQuantile regression or a bootstrap over the calibration set. Publish an interval, never a pointown code

Density map regression is the right choice for dense objects like wheat ears. Labelling a box around each of 400 ears in a frame is unaffordable; clicking a dot on each is not, and the dot is all a density model needs.

Cost efficiency

  • De-duplicate, or you will double count. At 70% overlap, every object appears in three frames. This bug is easy to ship and hard to notice, because the number is merely wrong rather than obviously broken. Validate against hand-counted quadrats every single season.
  • Publish a range, always. A point estimate that is 12% out destroys trust in the whole platform. A range that contains the answer builds it. This is a product decision that costs nothing.
  • Hand counts are your cheapest asset. Twenty quadrats per field costs an afternoon and calibrates everything. Build the capture form first.
  • Sell the decision, not the number. Nobody buys "an estimate". They buy "book six trailers instead of eight", "redrill these 3.1 ha", "the pickers come Thursday".