When can similar fields improve yield predictions?

A traceable correction for frozen yield models, scored forward in time.

How it works ↓Background, data, evaluation, method. Correct one field →Choose a correction, then see the harvest.

2,062 field-seasons3 crops4 countries2016–2024

Scroll Imagery © Esri, Maxar · illustrative

Across all fields

Without labels, the best correction depends on the crop.

–individual methods in the chart clearly help all three cropsIn-season · τ1–τ4

One choice per crop-year

Choosing from past seasons helps all three crops.

–crops with lower MSE, 95% intervals above 0In-season · τ1–τ4 · exploratory replay

Where the error lives

A field-level shift can reach most of the Base error.

–of Base MSE is a year or field offset, by cropIn-season · τ1–τ4

When it helps

Corn and soybean gain early, wheat late.

Year by year

Some years hurt.

–of fields made worse, depending on the methodIn-season · τ1–τ4

With same-season labels

A few labels go furthest in wheat.

–less wheat error when 1 in 10 fields is labelledIn-season · τ1–τ4

Limitations

Traceable corrections can still fail.

Wheat 2020 · field 37 · full season · chosen after the fact to show a failure

1Similarity can mislead.

Its closest past fields were over-predicted; this one was under-predicted.

2Abstaining is still a decision.

No policy kept a Base that was nearly unbiased on average. Knowing where to stop takes the harvest.

3Traceable is not fully explained.

We can replay the correction. Explanations of the selector remain unvalidated.

Evidence scope11 test crop-yearsOne benchmarkExamples chosen after the factOne Base per yearNo user study

Explorer

Now correct one field yourself.

Pick a correction, set λ, lock. Then see the harvest.

Open the Explorer →