What this research found
Satellite imagery can show where a field is under-performing but not why. Seven 2023 Sentinel-2 acquisitions over a 41-by-49 km stretch of central Iowa were combined with the USDA Cropland Data Layer to rank 1-km cells of near-pure corn or soybean by how far their vegetation indices sat below same-crop peers. The output is a 20-cell scouting list, each cell paired with a nearby normal control, plus a fully specified split-plot nitrogen-by-water field trial to test whether nutrient or moisture stress actually explains the anomalies.
- Filtering the 2023 Cropland Data Layer at 90% purity on a 1-km grid left 56 usable cells, 40 corn-pure and 16 soybean-pure, covering 522,739 native 10-metre crop pixels (374,098 corn and 148,641 soybean).
- Seasonal index profiles reproduced the expected corn-soybean offset. Corn peaked in July with a cell-mean peak NDVI of 0.635, persistence of 0.480, and peak NDRE of 0.562, while soybean peaked later at 0.615 NDVI, 0.422 persistence, and 0.495 NDRE, and stayed green longer.
- Corn anomalies are far clearer than soybean ones. The corn stress-index leaders run from -2.74 to -0.50, headed by cell 33 at -2.744, whereas the soybean leaders span only -1.48 to +0.08, which the report flags as a weak signal rather than presenting it as equivalent evidence.
- All 20 flagged cells were matched to unique same-crop controls, 18 of them inside the preferred 2 to 10 km band, with one at 10.3 km and one soybean cell at 27.7 km. Data-quality flags are carried alongside: 17 of the 20 cells are rated High overall uncertainty, and 4 soybean controls are themselves flagged cells.
- The confirmatory design is a randomised split-plot trial with water management (rainfed or irrigated) as the whole-plot factor and nitrogen (deficient, standard, luxury) as the sub-plot factor, 3 blocks per location and 360 plots of 12 by 24 metres, with water tested on 1 and 2 degrees of freedom and nitrogen and the interaction on 2 and 8.
How it was done
Seven Sentinel-2 Level-2A acquisitions of one tile between April and October 2023 were retrieved from a public catalogue, with red, red-edge, and near-infrared bands placed on a common 10-metre grid and cloud and shadow pixels masked using the scene classification layer. The national 2023 Cropland Data Layer was read tile-by-tile directly out of a 3.3 GB archive and reprojected by nearest neighbour to preserve categorical class codes, then used to find 1-km cells at least 90% corn or soybean. NDVI and NDRE were reduced per pixel to establishment, peak-vigour, persistence, and senescence metrics, averaged within each cell, and converted to z-scores against the same crop's own baseline so that corn and soybean phenology could not contaminate each other's scores. A minimum-distance Hungarian assignment then paired each flagged cell with a unique control, and a seeded randomisation produced the field layout, all documented in a 20-page report with 7 figures and 55 references.
Data sources
- Sentinel-2 Collection-1 Level-2A, tile T15TVG, 7 dates from 2023-04-06 to 2023-10-21 (Microsoft Planetary Computer STAC)
- USDA NASS Cropland Data Layer 2023, national 30-metre raster (corn 41.1% and soybean 32.7% of the study area)
- Central Iowa study area from -93.75 to -93.15 longitude and 41.85 to 42.22 latitude, gridded into 2,009 full 1-km cells
Limitations
Three of the seven acquisition dates fall on an orbit swath edge that images only about 64% of the area, leaving the establishment metric undefined for 20 of 56 cells and senescence for 19, and only 16 soybean cells passed the purity filter, so the soybean control pool could not supply unflagged matches. Vegetation indices detect spectral symptoms rather than causes, so a low index cannot separate nitrogen or water stress from pathogens, compaction, pests, or weeds, and a null result in the proposed trial would reject those two drivers rather than certify the field as healthy.
Figures from this analysis
How this research was produced
K-Dense Web planned and ran this agriculture investigation end to end — gathering the sources, carrying out the analysis, producing the figures, and drafting the report. The full session transcript, including every intermediate step, is available to view.
