Chrysostomou et al. (2026) Optimized spectral indices for global vegetation and water mapping using Sentinel-2

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-01-13
• Authors: Charalambos Chrysostomou, Stelios P. Neophytides, Michalis Mavrovouniotis, Diofantos G. Hadjimitsis
• DOI: 10.1038/s41598-025-34720-x

Research Groups

• ERATOSTHENES Centre of Excellence, Limassol, Cyprus

Short Summary

This study introduces two novel spectral indices, SRVI and SRWI, derived using a data-driven symbolic regression framework on Sentinel-2 data and ESA WorldCover labels. These indices demonstrate significantly improved separability for global vegetation and water mapping compared to established benchmarks, offering enhanced discrimination and reduced confusion across diverse biomes.

Objective

• To develop novel Symbolic Regression Vegetation Index (SRVI) and Symbolic Regression Water Index (SRWI) using a data-driven symbolic regression framework on Sentinel-2 imagery from a spectrally diverse Mediterranean site.
• To rigorously evaluate the performance and generalizability of SRVI and SRWI against established spectral indices across eleven independent, globally distributed regions.
• Hypothesis: Indices discovered through direct optimization on diverse spectra will yield statistically and practically significant gains under out-of-sample validation, demonstrating robust generalization across biomes and environmental conditions.

Study Configuration

• Spatial Scale: Training conducted over the Limassol district, Cyprus (406 km²). Validation performed across eleven globally distributed regions totaling 13,488 km². All data were processed at a 10 m spatial resolution.
• Temporal Scale: Data from 2021, with twelve single-date Sentinel-2 Level-2A acquisitions per study area (one per month, least-clouded scene), totaling 144 images. Monthly phenological trends were considered.

Methodology and Data

• Models used: Symbolic Regression (implemented with PySR) for autonomous discovery of mathematical expressions.
• Data sources:
  • Copernicus Sentinel-2 Level-2A surface reflectance products (visible, near-infrared, and shortwave infrared bands).
  • ESA WorldCover 2021 (global 10 m land cover map) as the reference standard for training and validation.

Main Results

• Two compact and interpretable indices were discovered:
  • Symbolic Regression Vegetation Index (SRVI): (2.0 * NIR - 3.0 * Red) / (1.0 * NIR + 1.0 * Red + 0.5 * Green + 0.5 * SWIR1)
  • Symbolic Regression Water Index (SRWI): ((Green + Blue) - (NIR + SWIR1)) / ((Green + Blue) + (NIR + SWIR1))
• Vegetation Mapping: SRVI significantly improved separability between vegetation and non-vegetation classes across eleven out-of-sample regions (e.g., Cohen’s d = 0.62 versus EVI, d = 0.42 versus MSAVI2; p < 0.01). It also showed consistent gains in discrimination among vegetation types (Tree cover, Shrubland, Grassland, Cropland) relative to SAVI, NDRE, and MSAVI2 (small-to-moderate effect sizes, p < 0.05), retaining sensitivity in high-biomass conditions where other indices saturate.
• Water Mapping: SRWI outperformed NDWI and MNDWI for water detection with significant improvements and moderate effect sizes (e.g., Cohen’s d = 0.60 versus NDWI). It produced a tighter, positive cluster for open water and consistently negative values for non-water classes, effectively suppressing urban and shadowed confounders. Performance was comparable to AWEI but with a simpler, sign-consistent formulation.
• Both indices demonstrated realistic seasonal behavior at the training site and robust generalization across diverse biomes and environmental conditions.

Contributions

• Introduces a novel data-driven methodology using symbolic regression for the autonomous discovery of compact and interpretable spectral indices, moving beyond traditional human-crafted designs.
• Develops two new indices, SRVI and SRWI, which exhibit enhanced statistical separability and practical utility for global vegetation and water mapping.
• Provides empirical evidence of robust generalization for these indices across diverse biomes and seasons through extensive out-of-sample validation.
• Offers practical gains for operational global mapping pipelines, improving accuracy for applications such as land-use change detection, crop condition screening, and flood mapping.

Funding

• European Union’s HORIZON Research and Innovation Programme: ‘EXCELSIOR’: ERATOSTHENES: Excellence Research Centre for Earth Surveillance and Space-Based Monitoring of the Environment H2020 Widespread Teaming project (Grant Agreement No 857510).
• Government of the Republic of Cyprus through the Directorate General for the European Programmes, Coordination and Development.
• Cyprus University of Technology.

Citation

@article{Chrysostomou2026Optimized,
  author = {Chrysostomou, Charalambos and Neophytides, Stelios P. and Mavrovouniotis, Michalis and Hadjimitsis, Diofantos G.},
  title = {Optimized spectral indices for global vegetation and water mapping using Sentinel-2},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-025-34720-x},
  url = {https://doi.org/10.1038/s41598-025-34720-x}
}