EUSFlux: Mapping Forest Carbon Uptake in the Eastern US from Upscaled Flux Tower Data

This repository provides annual maps of Net Ecosystem Productivity (NEP) for eastern US forests at 500 m resolution, 2003–2023. NEP is predicted by a Random Forest model trained on AmeriFlux and NEON eddy covariance observations and applied to gridded MODIS, solar-induced fluorescence (GOSIF), Daymet and ERA5-Land predictors. Predictions are made for 16-day periods, gap-filled, and summed (day-weighted) to annual NEP. From 51 candidate predictors (Table 2 of the paper), permutation importance across 100 repetitions of 5-fold spatial cross-validation retained 13; daylength was added, giving 14: forest type, NDVI, EVI, LSWI, the 30-day LSWI anomaly, NIRv, MOD09A1 surface reflectance Bands 2, 4 and 7, LAI, SIF, root-zone soil temperature, shortwave radiation and daylength. The Random Forest (100 trees, max_features = 0.33) was tuned by nested spatial cross-validation, and SHAP analysis was used to interpret it. The Zenodo repository includes: The preprocessed flux tower training dataset (NEE_flux_unh_sif_8.csv), containing 16-day site-level NEE observations and co-located predictor variables from AmeriFlux/NEON sites in eastern US forests; • An independent evaluation dataset of monthly FluxCom NEE estimates at seven held-out flux tower test sites (test_sites_monthly_fluxComNEE_2003_2021.csv; column NEE_gC_m2_d1); • The study area boundary shapefile (StudyArea.shp); • Python scripts, Slurm job scripts and one Jupyter notebook implementing the full workflow, in four folders: input data download and checks, predictor selection and model tuning, NEP prediction and annual forest maps, and the analyses and figures of the paper (run on Indiana University’s Big Red 200 cluster; file paths must be adapted); • Annual NEP GeoTIFFs (NEP_annual_YYYY.tif, 2003–2023) at 500 m resolution. NEP is reported in units of gC m⁻² year⁻¹ in annual rasters, with positive values indicating net carbon uptake and negative values indicating net carbon release. Pixels outside each year’s NLCD forest mask are NoData. The sign convention follows NEP = −NEE; flux tower data and intermediate prediction rasters are in NEE convention and are sign-flipped in the analysis scripts. A README provided within the deposit describes the contents, formats, column names, units, and provenance of all shared data files and code. Predictor variables and benchmark data are derived from the following external datasets, which are not included in this deposit and must be obtained separately: MODIS MOD09A1, MOD13A1, MOD15A2H, MCD12Q1 and land surface temperature products (NASA LP DAAC); GOSIF SIF (University of New Hampshire); Daymet v4 (ORNL DAAC); ERA5-Land (ECMWF / Copernicus Climate Change Service); AmeriFlux/NEON flux tower observations; FLUXCOM RS+METEO (Max Planck Institute for Biogeochemistry); NLCD land cover (USGS), 2003–2023; forest carbon removals from Harris et al. (2021) (Global Forest Watch); and a forest stand age raster for 2010.