Optimized spectral indices for global vegetation and water mapping using Sentinel-2.

Reliable mapping of vegetation and surface water from satellite imagery remains challenging, as common spectral indices can saturate at high biomass, show limited sensitivity across ecosystems, and confuse targets with soil, shadows, or built-up surfaces. We present two indices, the Symbolic Regression Vegetation Index (SRVI) and the Symbolic Regression Water Index (SRWI), discovered with a data-driven symbolic regression framework applied to Sentinel-2 Level-2A reflectance and guided by ESA WorldCover labels. Expressions were evolved from physically interpretable building blocks using non-linear combinations of visible, NIR, and SWIR bands. Indices were derived on a spectrally complex Mediterranean site and evaluated on eleven independent regions spanning diverse biomes. Performance was assessed with the Jeffries-Matusita distance, averaged across months to account for phenology, and compared against established vegetation indices (NDVI, EVI, SAVI, MSAVI2, NDRE) and water indices (NDWI, MNDWI, AWEI, TCW, WI2015). SRVI improves separability between vegetation and non-vegetation and shows higher discrimination among vegetation types relative to all benchmarks. SRWI yields more consistent water delineation with reduced confusion with built-up and shadowed surfaces, outperforming standard alternatives on the same datasets. Results indicate that symbolic regression can produce compact, interpretable indices that generalise across regions and seasons, offering practical gains for global vegetation and water mapping.