fieldworkAGRICULTURAL INTELLIGENCEMETHODS / A REPRODUCIBLE PUBLIC-DATA PROJECT
First, inspect what was aligned, flagged and withheld. Then see how plot-scale variation determines the replication needed for a small yield difference.
Measured machinery findings. Prospective power scenarios. No client trial, calibrated sensor correction or causal yield gain is claimed.
01 / MACHINERY DATA
2018 harvest positions are in WGS84 degrees; application points are in UTM 12N meters. We reproject the harvest into the application frame. The measurements remain intact.
Same coordinate frame, extent and yield scale within each harvest.
unique records flagged
15,302 of 30,133 QA-retained records
Nearest application point within 12 m. Coordinate reprojection does not establish GNSS accuracy.
| Diagnostic | 2018 | 2016 | Action |
|---|---|---|---|
| Coordinate-system mismatch | Degrees vs meters | Same source projection | Transform 2018 to EPSG:32612; preserve the unspecified 2016 datum. |
| Duplicate x / y / yield | 5 | 197 | Keep first in the analysis view; preserve the full ledger. |
| Non-positive yield or flow | 0 | 1,146 | Flag and exclude; no replacement values. |
| Speed below 1.5 mph | 342 | Not supplied | Screen 2018; no invented 2016 speed. |
| Moisture outside 8–22% | 247 | 82 | Flag outside the stated range. |
| Non-positive swath | 0 | 833 | Flag and exclude; source units retained. |
| Yield percentile tails | 626 | 1,501 | Screen outside central 98% of positive values. |
| Conflicting polygon rates | Not applicable | 592 raw cases | Withhold joins when overlap rate spread exceeds 5 source lb/ac. |
Quality reasons overlap. These screening choices can flag legitimate extremes; they are not all proven machine errors. All original rows remain inspectable.
COVERAGE / 2018
8 m matches 47.0% of QA-retained yield; 12 m matches 50.8%; 20 m matches 51.3%. Only 162 records are added from 12 to 20 m.
Proximity alone does not prove time alignment or randomized treatment assignment.
ACCURACY / UNRESOLVED
13,136 QA-retained harvest points are within 1 mm of an application coordinate. Numerical coincidence in source positions is not evidence of sub-millimeter GNSS accuracy.
Sensor delay, GPS drift, antenna offsets and yield calibration need supporting instrument evidence before a physical correction can be claimed.
The 2018 GeoJSON retains original WGS84 positions. No WGS84 layer is exported for 2016 because its datum is unspecified.
02 / FIELD-TRIAL REPLICATION
Two treatments, randomized within independent paired blocks. Each block contains one A plot and one B plot. GPS records within those plots are subsamples.
Two-sided alpha = 0.05. Exact paired-t power under normal, independent block differences. These are scenario inputs, not measured client variance.
COMPLETE PAIRED BLOCKS REQUIRED
52104 total plots / 52 per treatment
80.78% calculated power
Before plot or pair losses. Validate plot-scale variance and the planned design before choosing a field layout.
A BETTER VARIANCE ESTIMATE
Estimate the SD of paired treatment differences from comparable replicated trials or a pilot. If the two plot SDs equal σ and their within-block correlation is ρ, then SD(D) = σ√[2(1−ρ)].
The historical 20 m-cell residual SD is a proxy, not a validated estimate of independent trial-plot variation. Multi-farm, multi-season and spatially correlated designs need their own analysis.
IMPLEMENTATION / VERIFICATION
52 complete pairs with an assumed 10% pair loss means scheduling 58 blocks / 116 plots for expected retention. This is not a guarantee. Machine width, plot size, buffers and harvest calibration still need a protocol.
20,000 simulations per validation scenario agree with exact power within four Monte Carlo standard errors. Simulated zero-effect rejection was 4.88% against nominal 5%.
SOURCES / REPRODUCIBILITY
The 8-page report contains the full handling rules, match sensitivity, assumptions, replication table and references. The project includes source hashes, audit rows, scripts and original frozen files.