FLUXtrapolation

FLUXtrapolation is a benchmark dataset designed to evaluate and advance methods for ecosystem flux prediction under realistic, progressively challenging forms of extrapolation. It addresses a critical gap in existing benchmarks by focusing on flux upscaling, a form of domain generalization characterized by spatiotemporal heterogeneity, as well as shifts in both input distributions and input–target relationships. The dataset comprises of hourly observations from 207 FLUXNET eddy-covariance sites between 2015 and 2022, spanning a wide range of climates, biomes, and geographic regions. It includes three key ecosystem flux targets: gross primary productivity (GPP), net ecosystem exchange (NEE), and evapotranspiration (ET), as well as 13 input covariates derived from meteorological measurements, site characteristics (e.g., IGBP vegetation type), and satellite remote sensing (e.g., EVI, NDWI), all processed through the FLUXCOM-X framework. FLUXtrapolation provides a standardized, scientifically grounded platform for diagnosing model failures, understanding extrapolation challenges, and developing more reliable methods for ecosystem monitoring and climate modeling.