Satellite Data Cannot Remove Basis Risk from Crop Insurance

The satellite image is usually presented as the end of an argument. The field is visible. The crop is visible. The drought is visible.

Therefore the loss is visible.

That last step is where the product quietly changes shape.

A satellite image can measure a field. It cannot measure the gap between an index payout and the farmer’s actual financial injury.

That gap is basis risk.

Better imagery sharpens our signal.

It cannot make the signal and the loss the same thing. Which matters here, because Canadian crop insurance is already several deliberately different measurement systems. A provincial AgriInsurance contract asks whether production for a crop across the farm fell below its guarantee. A private hail policy asks what percentage of liability was destroyed on nominated acres. A parametric cover asks whether a selected number crossed a selected trigger.

All three can be sensible.

None of them is interchangeable because all three use data about a field. The technology question isn’t whether remote sensing is accurate enough to remove adjustment. It is whether the programme says, in advance, what the signal decides, what it merely supports, and what happens when the farm and the signal disagree.

A clearer image is not a loss adjustment

Take the Saskatchewan grain-farm flow.

The producer renews AgriInsurance by the end of March and picks coverage, typically half to four-fifths of the farm’s long-term average yield for each crop. On 1,200 acres of durum, a long-term average of 38 bushels an acre at 70% coverage gives a 26.6 bushel guarantee.

That is 31,920 bushels across the crop.

At the published insured price, roughly $335,000 of liability.

The important word is across.

The guarantee we sold is not a promise about any field. It is a farm-level bushel total for that crop.

The seeded-acreage report is due by late June.

It declares what went into each field, on which land, on what seeding date.

It fixes the base we insure.

Unreported acres are uninsured.

Crops seeded after the final seeding date carry reduced coverage, or none. Satellite acreage and vegetation checks help verify that declaration. They are useful precisely because they make a mismatch visible without sending an adjuster down every road in the province. They do not change the contract the declaration supports. On 12 July, hail strips one quarter section of durum. The private hail insurer assigns an adjuster, who works the affected legal land description, counts surviving stems, examines bruised nodes and shattered heads, and assigns a percentage loss. At $300 an acre of liability, 160 acres, and a 55% assessed loss, the payment is $26,400.

A spot-loss settlement.

Damage on those acres, on that afternoon.

At harvest the public programme measures something else entirely. By mid-November the producer reports harvested production, using elevator settlement statements, scale tickets, and measured bin volumes. The agency sets the crop total against the guarantee, and audits a sample of farms by measuring bins. The undamaged durum yields 34 bushels an acre on 1,040 acres.

The hailed quarter produces 17.

Total production is 38,080 bushels against a 31,920-bushel guarantee.

The hail policy pays.

AgriInsurance does not.

Neither system failed.

They answered different questions we asked separately.

Which is a useful discipline for a remote-data programme.

Before we connect an image to a payment, ask what it is doing.

Supporting a seeded-acreage check? Selecting a field for inspection? Estimating an area yield? Determining a cash payment?

Those aren’t stages of one maturity model.

They are different authorities.

The index becomes the product

Parametric wording does not move loss uncertainty out of the contract.

It moves it forward in time.

For drought cover the index can be rainfall or vegetation data: CHIRPS, ERA5 reanalysis, MODIS-derived NDVI.

Our product definition names the source, the grid cell or station list, the measurement window, and the averaging rule.

It also says whether the first published value or a later restated one governs, when revisions stop changing settlement, and what happens if the source stops publishing.

Those aren’t technical implementation details.

They are the insurance product.

CHIRPS is revealing, because its satellite-and-gauge blend can operate on a grid cell of roughly five kilometres.

A scheme can calculate dekadal rainfall for germination, vegetative, and grain-fill windows, then pay when the cell falls below trigger.

Fast, and lean for us to run.

Also an average over farms that may have had very different rain.

The farm that loses its crop while a neighbour two kilometres away gets paid does not experience that as a resolution problem.

It experiences a promise we made and could not keep.

The farmer wasn’t wrong to expect a crop-loss product.

The product was wrong to imply its cell was a farm.

Satellite data improves the observation.

It does not remove the distinction between a grid and a field.

Nor does it remove metric risk.

Rainfall is not yield. Vegetation is not margin.

A well-watered crop can still suffer a financial injury.

A weak vegetation signal can occur with no payable shortfall under a farm-wide yield guarantee.

There is temporal risk too.

The wording picks the window in which a reading counts.

And design risk: a binary trigger, a stepped payout table, and a linear function make different promises at the boundary.

Attachment is where we start paying.

Exhaustion is where the full limit is payable.

An index can be technically precise and contractually coarse.

Here is the limit I would state.

We can say remote sensing cuts the cost and time of verification.

We should not say it eliminated basis risk, unless we also eliminated the gap between the contracted reading and the actual loss.

We did not.

Canadian crop insurance has a different centre of gravity

The Canadian public programme is not a generic index template waiting for a satellite feed.

The Farm Income Protection Act lets the federal Crown make agreements with provinces.

Under the Sustainable Canadian Agricultural Partnership framework, the provincial Crown corporation issues the actual contract.

SCIC in Saskatchewan. AFSC in Alberta. Agricorp in Ontario. MASC in Manitoba. La Financière agricole in Quebec.

A producer pays premium, meets deadlines, supplies records, and can be denied a claim.

Public funding does not make that an informal government payment.

The rules reflect the structure.

We certify the plan actuarially sound.

Coverage uses the producer’s individual production history, blending farm history with area average where records are short.

The thirty per cent threshold in several programme rules is a trade-law artefact from the WTO exempt-support annex.

It is not a satellite-performance target.

Indemnities also count as programme income for AgriStability margin calculations.

So we cannot lift a generic parametric template into a Canadian crop programme and call it modernisation.

It could replace a farm-level production test with a different promise.

It could change the evidence path, shift the distribution of shortfalls, and produce a dispute the appeal machinery was never built to hear.

That may still be a valid product decision.

It has to be named as one.

Private hail is the healthier comparison.

It has always been a deliberately local answer to a deliberately local loss.

It sits beside AgriInsurance because it pays a different measure of harm.

The public programme absorbs a field loss when the rest of the farm performs.

The hail policy never pretended otherwise.

A parametric layer can be equally honest, positioned as a rapid, bounded payment for a named cash-flow need.

Not as a quiet replacement for indemnity.

Speed needs an exception owner

The attraction of parametric settlement is real.

Once the governing index publishes, a calculation agent applies the stated formula and determines the payment.

No field adjustment sits between publication and money.

CCRIF SPC states that its payouts go out within a fortnight of an event, because the structure does not wait for loss verification.

But the absence of adjustment is not the absence of process.

We still need the data source, the index publication, the payout function, and a calculation agent.

It needs an index-cessation clause: a fallback source, or authority for the agent to determine the value in good faith and in a commercially reasonable manner.

Otherwise an outage creates an unresolvable gap.

Commercial parametric cover commonly adds a proof-of-loss attestation.

The insured declares an insurable interest and an actual loss at least as large as the payment.

Not a conventional adjustment, and it can take days rather than months.

It does mean the payment shrinks when the actual loss is smaller.

The upside of basis risk goes.

The downside stays.

For a Canadian crop programme, the familiar evidence path stays visible wherever the promise depends on actual yield.

A producer gives notice before grazing, baling, or working down a damaged crop.

The agency appraises it and leaves standing check strips.

Production records and bin measurement stay available at harvest.

Destroy the crop before appraisal and the claim on those acres is forfeited.

The evidence is gone.

Remote sensing makes our route better.

It highlights a field needing inspection, tests whether reported acreage is plausible, and sends scarce adjusting capacity to the exception.

It should not decide that a visible pattern makes check strips, records, and the right to challenge an appraisal unnecessary.

A clean dashboard is not evidence where the product promises an indemnity.

Say which disagreement we are having

The board hears that a satellite platform improves speed, coverage, and data quality.

We need a narrower question.

What does the platform have authority to decide?

The answer produces a controlled record.

Per crop and peril: insured unit, seeded-acreage declaration, data source, grid or field boundary, measurement window, index version, publication rule, trigger, attachment, exhaustion, payout schedule, disruption fallback, calculation-agent determination, production evidence, and the route for an objection.

That record separates four disagreements we otherwise label one vague “data-quality issue.”

The reading may not represent the field.

The reading may represent the field and not the loss.

The loss may fall outside the agreed window.

Or the payout table may put two nearly identical readings on opposite sides of a step.

Each needs a different owner.

The board framing is product honesty.

Quantify which losses the signal catches, which it misses, and who owns the exception when index and farm diverge.

Don’t put that last question into a queue called manual review.

The farmer who has lost a crop will not accept an explanation of pixel resolution as a settlement.

Use remote data to make reporting and triage faster.

Keep the yield, hail, and index promises separate.

The better the signal gets, the more important it is to say what it still does not indemnify.