Accès ouvert · CC BY

Systematic vegetation index discovery for phenology-based crop mapping

Article scientifique 2026 Sciences de la Terre et de l'environnement › Sciences de l'environnement (OCDE 1.5) Anglais

Résumé

Introduction The Systematic Vegetation Index Discovery (SVID) framework generates and evaluates large numbers of vegetation indices (VIs) by combining spectral bands with basic arithmetic operations. We evaluated whether systematically discovered indices could improve agricultural land-cover classification compared with conventional indices such as NDVI while reducing the dimensionality of multitemporal spectral data. Methods We applied SVID to a 323.9 ha agricultural landscape in Andong, South Korea, containing rice paddies, pepper fields, and plastic-covered greenhouses, using 32 cloud‐free Sentinel‐2A images acquired during 2022. Of 1,328 candidate expressions, 1,144 produced numerically valid rasters. A separate Random Forest model was trained for each index using its 32‐date temporal profile and evaluated against 500 independent reference points withheld from training. Before ranking, 388 expressions were identified as algebraically identical to another expression, and a further 355 were strictly monotone transformations and therefore indistinguishable to a rank‐based classifier. Results The 1,144 indices consequently represented 401 distinct features. Several near‐infrared‐to‐visible ratio families outperformed NDVI. The best‐performing index, (NIR – Red)/(Green × Blue), achieved 97.0% overall accuracy and a Kappa coefficient of 0.873, compared with 94.4% and 0.734 for NDVI; this difference was significant by McNemar's test (p = 0.002). The improvement was concentrated in the greenhouse class, whose producer's accuracy increased from 58.6% to 84.5%. A model using all four raw bands across all dates (128 features) achieved 96.0% accuracy and was statistically indistinguishable from the best single index, demonstrating a four‐fold reduction in dimensionality without a detectable loss in accuracy. Rank positions were nevertheless unstable: ten of the 401 features were statistically tied with the best, and mathematically identical formulations differed by up to 3.4 percentage points because of finite‐precision effects. Discussion SVID can identify compact spectral representations that perform as well as or better than conventional indices and full multiband temporal models. However, the instability among top‐ranked indices suggests that results should be reported as index families with confidence intervals rather than as a single winner. Limitations include the use of one site, one year, and four spectral bands; broader testing across sites, years, and spectral configurations is needed.

Citer ce document

Rahimi, E., & Jung, C. (2026). Systematic vegetation index discovery for phenology-based crop mapping. Frontiers in Remote Sensing. https://doi.org/10.3389/frsen.2026.1911652

Exporter : BibTeX · RIS (Zotero, Mendeley, EndNote)

Accès au document

Texte intégral archivé sur SPHAERO, en accès ouvert : lisible gratuitement en ligne après connexion. Le document reste aussi disponible à sa source. Se connecter

Voir l'article sur le site de la revue

Licence et provenance

Licence : CC BY

Notice moissonnée depuis OpenAlex le 19/10/2026. Le document reste hébergé par sa source.
Voir le document à la source →

Statistiques

Consultations (hors robots) : 12

Téléchargements (hors robots) : 0

Méthode de comptage