From combine records to a defensible interpretation. Explore two wheat harvests, inspect the data, and see how uncertainty changes the story.
See where the combine travelled, what it recorded, and where application data can support an analysis. Select a point to inspect its source values.
Screening rules are transparent analytical choices. Flags identify records for review; they do not establish instrument errors or repair calibration.
See the evidence boundary ↗Raw averages can hide spatial patterns. Compare a rate-only model with a model that also accounts for location, then inspect how screening choices affect the result.
Curve shown over the middle 80% of observed cell rates. Each dot is an equally weighted cell, not an independent treatment replicate.
Contrast between the 25th and 75th percentile of matched cell rates. Interval resamples 100 m spatial blocks; it is conditional on the selected model.
Profiles can change both the retained records and the observed rate contrast. This is a sensitivity comparison, not three independent experiments.
Residual semivariance by distance. Remaining spatial dependence can make intervals too narrow. Block holdout measures within-field prediction only.
Model: recorded yield = intercept + recorded rate + rate² + x + y + x² + xy + y². Ordinary least squares on 20 m cell medians, with at least three matched harvest observations per cell.
Uncertainty: 400 resamples of 100 m spatial blocks, using a fixed seed. The blocks are analytical groups, not randomized plots. Spatial dependence beyond block boundaries may understate uncertainty.
Interpretation: the randomized trial layout and independent experimental-unit identifiers have not been established. This model describes a spatially adjusted association. A trial-specific mixed or spatial covariance model requires the original design and treatment assignments.
Explore how independent replication and field variability affect the chance of detecting a future yield difference.
SIMULATION · PLANNING SCENARIOCalibrated to the 2018 balanced model’s cell residual SD, used only as a planning proxy. Assumes independent paired-plot blocks, within-pair correlation 0.5, and a two-sided 5% test.
—
Each replicate is one independently randomized block with two plots. More GPS readings do not create more treatment replicates. Confirm actual plot variance before prescribing sample size.
Every map starts with a published file. Every estimate has a reproducible method. Every limitation stays attached to the result.
Public example data from Montana State University’s On-Farm Precision Experiments project. Original archives and SHA-256 hashes are included.
View the source repository ↗ae70c5d8ad4f…No soil, weather, satellite imagery or randomized treatment layout is invented. No machinery delay correction is claimed without validated timing.
The 2018 rate documentation conflicts on product versus nitrogen mass. We retain the recorded rate and do not convert it to nitrogen.
Take the current batch, screening choices and model results into an agronomist-friendly report, with methods and limits included.