FluxCubes30: A 30-m, 27-year (1999–2025) global flux-tower-centered GPP spatial cutout dataset based on optimal footprint and cross-calibrated Landsat observations

As Earth system science advances into the deep-learning era, geoscience foundation models require high-resolution training samples with explicit spatial context. To complement point-scale eddy-covariance ground observations, we present FluxCubes30, a global flux-tower-centered, high-resolution (30 m), 27-year (1999–2025) gross primary productivity (GPP) spatial cutout dataset. This dataset encompasses 404 core sites across major regional flux networks. It was generated by harmonizing Landsat 7, 8, and 9 observations through a cross-calibration algorithm to drive an ecosystem light-use-efficiency (eLUE) model. The model is forced by Landsat-based Enhanced Vegetation Index (EVI) and top-of-atmosphere photosynthetically active radiation (PAR_TOA). The production pipeline incorporates a dynamic footprint optimization algorithm and propagates pixel-level 95% confidence uncertainty for each 30 m grid cell. As an AI-ready spatiotemporal benchmark, FluxCubes30 reconstructs decadal carbon-sink histories for sites with short operational records, captures micro-refugia following extreme disturbances, and preserves agricultural irrigation signals that are typically smoothed out by coarse-resolution products. Dataset Characteristics Temporal Coverage: 1999–2025. Spatial Unit: 7 km x 7 km site-centered cutouts. Spatial Resolution: 30 m. Coordinate Reference System: WGS84 geographic coordinates, EPSG:4326. Main Output Variables: EVI, eLUE-modeled GPP (g C m-2 d-1), and Propagated GPP Uncertainty. File Organization The root directory contains the master index, documentation, and compressed archives organized by regional flux networks (e.g., JapanFlux.rar, Demo_Sites.rar). For networks that contains a substantial number of sites (e.g., AmeriFlux), the network is further divided into sub-files names as NETWORK + (SITE_NAME_INITIAL_LETTER).zip (e.g., AmeriFlux (A).zip). IndexMeta.csv: The master metadata index used for site lookup, model parameters, and relative file paths. FluxCubes30_User_Guide_v1.0.pdf: The comprehensive product user guide. DataRead_Example.R: An R script demonstrating how to batch read the metadata, CSVs, and multi-band GeoTIFFs. Within each individual site folder (e.g., AU-Ade): <Site_ID>_timeseries.csv: Site-level time-series containing Landsat observation-level EVI, PAR, GPP_EC, GPP_eLUE, and uncertainty bounds. <Site_ID>.png: A four-panel quick-look visualization. spatial_cutouts/: A directory containing three multi-band GeoTIFF cubes (EVI, GPP_eLUE, GPP_uncertainty) and a band_date_index.csv mapping each GeoTIFF band to its specific observation date. Usage Notes & Scaling Factors To optimize storage, all dynamic spatial cubes are stored as signed 16-bit integers. The NoData value across all rasters is -32768. Users must apply the following scaling factors to recover physical values before analysis: EVI: Divide raw values by 10000. GPP_eLUE: Divide raw values by 100 (Unit: g C m-2 d-1). GPP_uncertainty: Divide raw values by 100 (Unit: g C m-2 d-1). For a detailed description of the model parameterization, quality control, and data reading examples, please refer to the attached User Guide.