Crop-specific vegetation indices and phenological windows for spectral separability of 46 United States crops
Vegetation indices designed for greenness are routinely applied to crop discrimination, yet searches for better indices rarely control for redundant candidates, selection bias or spatial dependence. We quantified one-versus-rest spectral separability for 46 crops on 16,647 parcels across the contiguous United States in 2019, 2021 and 2023, using 36 dekadal Sentinel-2 composites. A generative grammar produced 985 index expressions, of which 563 were distinguishable to rank-based methods; each was scored at every dekad against spatially matched neighboring parcels, and leading candidates were re-evaluated with selection confined to training folds. Optima were largely crop-specific (38 different best indices for 46 crops), with median held-out separability of 0.846 across years. Once candidate-family size was equalized, no index family dominated. Held-out separability fell from 0.872 in training to 0.835 when years were withheld and to 0.734 when spatial blocks were withheld, a gap that persisted after removing test parcels near training parcels. Thinning clear observations left a fixed index essentially unchanged but degraded re-selection; nevertheless, re-selection within a withheld region outperformed, on average, indices transferred from other regions at every observation density tested. Index atlases should therefore report sets of near-equivalent indices, calendar windows and the domain over which separability was validated.
Authors
- Chuleui Jung (ORCID: https://orcid.org/0000-0001-8134-9279)
- Ehsan Rahimi (ORCID: https://orcid.org/0000-0002-8401-472X)
Institutions
- Gyeongguk National University (KR)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1038/s41598-026-79394-1
- Primary Topic
- Remote Sensing in Agriculture
- Type
- article
- Field-Weighted Citation Impact
- 0.00