Source-linked AI summary
The FLUXCOM ensemble of global land-atmosphere energy fluxes
Martin Jung, Sujan Koirala, Ulrich Weber, Kazuhito Ichii, Fabian Gans, Gustau-Camps-Valls, Dario Papale, Christopher Schwalm, Gianluca Tramontana, Markus Reichstein
TL;DR
Global land-atmosphere energy fluxes remain poorly constrained because tower measurements are sparse and uncertain. The paper merges tower, remote-sensing, and meteorological data with machine learning to produce an uncertainty-aware FLUXCOM ensemble, finding robust flux-pattern estimation across methods and setups.
Problem
Global land-atmosphere energy fluxes have large uncertainties, while tower observations are sparse and affected by unresolved energy-balance closure gaps.
Method
Machine learning merges FLUXNET tower measurements with remote sensing and meteorological data across two setups, correction approaches, and forcing choices to generate an ensemble.
Results
147 global gridded products were generated, with negligible performance differences across machine-learning techniques and between RS and RS+METEO setups.
Takeaways & Limitations
The FLUXCOM ensemble provides global flux products for quantifying land-atmosphere interactions and benchmarking land-surface-model simulations.
Takeaways & Limitations
Interannual variations are expected to be more uncertain than mean spatial and seasonal flux patterns and may be too small in magnitude.
Abstract
from arXiv · showhide
Although a key driver of Earth's climate system, global land-atmosphere energy fluxes are poorly constrained. Here we use machine learning to merge energy flux measurements from FLUXNET eddy covariance towers with remote sensing and meteorological data to estimate net radiation, latent and sensible heat and their uncertainties. The resulting FLUXCOM database comprises 147 global gridded products in two setups: (1) 0.0833$°$ resolution using MODIS remote sensing data (RS) and (2) 0.5$°$ resolution using remote sensing and meteorological data (RS+METEO). Within each setup we use a full factorial design across machine learning methods, forcing datasets and energy balance closure corrections. For RS and RS+METEO setups respectively, we estimate 2001-2013 global (${\pm}$ 1 standard deviation) net radiation as 75.8${\pm}$1.4 ${W\ m^{-2}}$ and 77.6${\pm}$2 ${W\ m^{-2}}$, sensible heat as 33${\pm}$4 ${W\ m^{-2}}$ and 36${\pm}$5 ${W\ m^{-2}}$, and evapotranspiration as 75.6${\pm}$10 ${\times}$ 10$^3$ ${km^3\ yr^{-1}}$ and 76${\pm}$6 ${\times}$ 10$^3$ ${km^3\ yr^{-1}}$. FLUXCOM products are suitable to quantify global land-atmosphere interactions and benchmark land surface model simulations.
Background & Summary
FLUXCOM addresses substantial uncertainty in global land-atmosphere energy fluxes by combining FLUXNET tower measurements with remote sensing, meteorological data, and machine learning. It provides an ensemble of global gridded flux products designed to characterize uncertainty and support comparisons with land surface models.
- Background: Global land surface models show large uncertainty in the magnitude and spatial pattern of land-atmosphere energy fluxes.Unevenly spaced FLUXNET tower measurements and approximately 20% energy-balance nonclosure complicate direct model comparisons and may introduce systematic bias.
- Motivation: FLUXCOM targets uncertainty from machine-learning algorithms, predictor variables, climate forcing data, and energy-balance closure.The initiative builds on earlier data-driven global flux products while explicitly examining multiple uncertainty sources in empirical upscaling.
- Validation: Cross-validation showed good tower-site performance for latent and sensible heat, especially net radiation, including seasonality and between-site mean fluxes.These predictions showed more skill than carbon-flux estimates, with negligible differences between tested alternatives reported in the supplied passage.
- Product design: The ensemble combines nine RS methods and three RS+METEO methods with energy-balance correction variants and, for RS+METEO, four climate-forcing products.RS products cover 2001-2015 at 8-day and 0.0833° resolution, while RS+METEO products use daily temporal and 0.5° spatial resolution.
Methods · Training of machine learning algorithms · RS RS+METEO
FLUXCOM machine-learning algorithms were trained on quality-screened observations from 224 flux towers, using setup-specific datasets for the RS and RS+METEO configurations. The training retained all available data points and followed the setup specifications summarized in Table 1.
- Methods: 224 flux tower sites supplied observations for training the machine-learning methods.The data followed specifications in Table 1 and previously detailed procedures.
- Training of machine learning algorithms: 20% was the maximum share of gap-filled half-hourly data allowed in a daily value during quality screening.Screening also required empirical consistency of energy fluxes and visual inspection.
- Training of machine learning algorithms: The same valid data points were used for net radiation and the other trained flux quantities.The passage also references potential shortwave radiation and individual time-step and location data.
- Training of machine learning algorithms: Each machine-learning method used the same training dataset within the RS or RS+METEO setup.This preserved setup-specific consistency across methods.
- Training of machine learning algorithms: All available data points were used for training, rather than only 90% of sites as in the cross-validation analysis.The contrast applies to training within each RS or RS+METEO setup.
- RS RS+METEO: Table 1 specifies the FLUXCOM RS and RS+METEO setups for energy fluxes.The table includes predictor-related acronyms such as EVI, fAPAR, LAI, LSTDay, LSTNight, MIR(1), NDVI, NDWI, and PFT.
Correction for energy balance non-closure at FLUXNET sites prior to training · Global products of predictor variables for RS products
FLUXCOM addressed FLUXNET energy-balance non-closure through three correction approaches and generated gridded predictors from MODIS remote-sensing products. The remote-sensing workflow filled poor-quality observations and aggregated or estimated values to create continuous grid-cell time series.
- Correction for energy balance non-closure at FLUXNET sites prior to training: Three correction approaches represented different hypotheses about the primary cause of the energy-balance closure gap at FLUXNET sites.The general correction was x_LE*LE+x_H*H=R_n-G, with x_LE and x_H correcting latent and sensible heat, respectively.
- Correction for energy balance non-closure at FLUXNET sites prior to training: The correction framework imposed x_LE*LE+x_H*H=R_n-G before machine-learning training.x_LE and x_H were the correction factors for latent and sensible heat, respectively.
- Correction for energy balance non-closure at FLUXNET sites prior to training: The Bowen ratio correction was identified as perhaps the most widely used approach among the correction methods.The passage states that this approach assumes a particular explanation for the closure gap, but the supplied text does not complete that assumption.
- Global products of predictor variables for RS products: MODIS land products from collection 5 supplied the remote-sensing inputs for FLUXCOM.The trained algorithms required only spatio-temporal grids of input data to produce spatio-temporal energy-flux grids.
- Global products of predictor variables for RS products: The MODIS predictors included daytime and nighttime land surface temperature, land cover, fPAR, and BRDF products.The supplied passage identifies the land-surface-temperature, land-cover, and fPAR product codes, while the BRDF entry is truncated.
- Global products of predictor variables for RS products: 25% of 1 km pixels within a 0.0833° grid cell had to have good quality for their mean to be used.Otherwise, the value was estimated using the local mean seasonal cycle; poor-quality data were filled to create continuous time series.
Solar Radiation · Global products of predictor variables for RS+METEO products
The section describes radiation and MODIS-based predictor inputs for FLUXCOM, including JASMES shortwave radiation, PFT-tiled land-surface variables, interpolated seasonal cycles, land-cover fractions, and four meteorological forcing datasets.
- Solar Radiation: 2001-2015: The RS product used incoming surface shortwave radiation from JAXA’s JASMES product, derived from Terra MODIS with a simple radiative transfer model.The product was previously evaluated at three EC sites in Asia and 20 EC sites in Alaska, showing good agreement with observations.
- Solar Radiation: The JASMES radiation products were previously evaluated against observations at three EC sites in Asia and 20 EC sites in Alaska.The passage reports good agreement with observations.
- Global products of predictor variables for RS+METEO products: 0.5°: MODIS-based remotely sensed land-surface variables were tiled by plant functional type, creating grids containing mean values per PFT and time step.The section refers to Table 1 and Tramontana et al. for details.
- Global products of predictor variables for RS+METEO products: Daily MODIS seasonal cycles for each RS+METEO grid cell were computed by linearly interpolating a temporally smoothed 8-daily mean seasonal cycle.This procedure supplied daily seasonal-cycle predictors for each grid cell.
- Global products of predictor variables for RS+METEO products: Land-cover fractions for the RS+METEO predictors used the same product and approach as in the RS product.The passage links the land-cover treatment across both setups.
Generation of global products (Prediction)
The trained machine-learning models generated RS products every 8 days at 0.0833° resolution and RS+METEO products daily at 0.5° resolution, with the latter aggregated across plant functional types.
- Generation of global products (Prediction): RS products were generated every 8 days at 0.0833° resolution by applying trained machine-learning models to gridded predictor fields.The models were applied at each 8-daily time step.
- Generation of global products (Prediction): RS+METEO products were generated daily at 0.5° resolution separately for each plant functional type.A weighted mean over plant functional type fractions was then obtained for each gridcell and time step.
- Generation of global products (Prediction): RS+METEO gridcell values were obtained by weighting plant functional type outputs by their fractions.The weighting was performed for each gridcell and time step.
Spatial and temporal aggregation of FLUXCOM-RS products
FLUXCOM-RS products were aggregated from their original 0.0833° and 8-daily resolution into monthly products at 0.0833° and 0.5° resolution. Monthly values were computed from interpolated daily data, while spatial aggregation used means of valid observations within each 0.5° cell.
- Spatial and temporal aggregation of FLUXCOM-RS products: Monthly FLUXCOM-RS products were derived at 0.0833° and 0.5° resolution from the original 0.0833° products with an 8-daily time step.The aggregation was intended to facilitate broader reuse of the products.
- Spatial and temporal aggregation of FLUXCOM-RS products: Monthly averages were calculated after linearly interpolating the 8-daily data into daily data.
- Spatial and temporal aggregation of FLUXCOM-RS products: 0.5° spatial values were calculated as the mean of non-missing data points within each 0.5° cell.A 0.5° grid also records the number of valid data points from the original RS product per cell.
Ensemble estimates
FLUXCOM ensemble products pool runs across machine-learning methods, forcing datasets, and energy-balance corrections. Monthly grid-cell estimates use ensemble medians, with median absolute deviation representing robust ensemble spread and uncertainty.
- RS ensembles: 27 ensemble members for LE and H and 9 for Rn were generated in the RS setup by pooling runs across 9 machine-learning methods and correction variants.RS products were generated at 0.0833° and 0.5° spatial resolutions; three energy-balance correction variants applied only to LE and H.
- RS+METEO ensembles: 9 ensemble members for LE and H and 3 for Rn were generated for each climate-forcing-specific RS+METEO ensemble.These ensembles pooled runs across three machine-learning methods and three energy-balance correction variants for LE and H.
- RS+METEO ensembles: 36 ensemble members for LE and H and 12 for Rn were generated for the overall RS+METEO ensemble.The overall ensemble pooled runs across climate-forcing datasets, machine-learning methods, and energy-balance correction variants.
- Ensemble estimates: Median monthly fluxes for each grid cell define ensemble estimates, while median absolute deviation provides a robust estimate of ensemble spread and uncertainty.For RS, spread captures uncertainty from machine-learning method choice and the lack of energy-balance closure in FLUXNET data.
Cross-consistency checks with the state-of-the-art estimates
FLUXCOM fluxes were cross-checked against established evaporation and radiation products using common spatial and temporal comparison protocols. Sensible heat and net radiation were additionally adjusted to account for non-vegetated land areas.
- Latent heat comparisons: FLUXCOM LE estimates were compared with MTE10, GLEAM v3.1a, and LandFlux-EVAL using spatial patterns and monthly continental-mean time series.The comparison included both the RS and RS+METEO ensembles.
- Net radiation comparisons: FLUXCOM Rn estimates were compared with the CERES SYN1d Ed4A and SRB release 3.1 satellite-based products.The original CERES and SRB 3-hourly data were aggregated to monthly means for comparison.
- Comparison protocol: Comparisons used the common 2001-2005 period, a common vegetated-land mask, and grid cells with at least 80% land fraction.Mean annual spatial fluxes were assessed with Pearson correlation, total least-squares fits, and density scatter plots.
- Global-area adjustment: FLUXCOM sensible heat and net radiation were scaled to include non-vegetated areas, chiefly cold and hot deserts, for comparison with global literature values.Hot-desert estimates used CERES net radiation and GPCP precipitation, while cold-desert values were based on reanalysis.
- Global-area adjustment: 5.9356 MJ m-2 day-1 for Rn and 5.8264 MJ m-2 day-1 for H were obtained for hot deserts during 2001-2010.These averages applied where hot desert exceeded 50% of a grid cell.
Code availability
Python code for synthesizing results and generating Figures 2–8 is publicly available, while MATLAB code for generating flux products and ensemble estimates can be requested for reproducibility.
- Code availability: Python code for synthesizing results and generating Figures 2 to 8 is available through a public repository.Repository: https://git.bgc-jena.mpg.de/skoirala/fluxcom_ef_figures
- Code availability: MATLAB code for generating flux products and ensemble estimates is available on request to Martin Jung for reproducibility.Contact: mjung@bgc-jena.mpg.de
- Code availability: The FLUXCOM initiative produced complex, large-scale code customized to high-performance computing.The passage attributes this complexity to the initiative’s collaborative nature and demanding computing requirements.
Data Records
FLUXCOM data files are distributed in netCDF-4 format using filenames that encode energy flux, setup, correction, machine-learning method, meteorological forcing, spatial resolution, temporal resolution, and year. The records include RS and RS+METEO products at monthly resolution, with spatial resolutions and ensemble/member options identified in the naming convention.
- Setup and product type: RS and RS_METEO identify the two upscaling setups, while TYPE distinguishes ensemble from member products.RS uses spatial-resolution codes 720_360 or 4320_2160; RS+METEO uses meteorological-data identifiers.
- File format and naming: Files use netCDF-4 format and follow a structured naming convention encoding product characteristics and year.The filename pattern is <EF>.<SETUP>.<EBC>.<MLM>.<METEO>.<sRESO>.<tRESO>.<YYYY>.nc.
- Correction metadata: Energy-balance correction codes identify ensembles, uncorrected fluxes, Bowen-ratio corrections, and residual corrections.Rn always uses NONE because it was never corrected.
- Machine-learning metadata: Machine-learning codes identify all-method ensembles or individual methods, including ANN, MARS, RF, and other methods depending on setup.RS+METEO products use ANN, MARS, or RF codes, whereas RS products include additional method identifiers.
- Resolution and year: Spatial codes 720_360 and 4320_2160 represent 0.5°x0.5° and 0.0833°x0.0833° grids, respectively, while all files use monthly temporal resolution.The year field identifies the data year.
Usage Notes
FLUXCOM usage guidance favors spatial patterns of mean annual and seasonal fluxes, with setup choice depending on the comparison or budget application. Interannual anomalies, land-area accounting, and ensemble uncertainty require specific normalization, scaling, or covariance assumptions.
- Comparison guidance: Use spatial patterns of mean annual and seasonal fluxes for cross-consistency analysis and evaluation of land surface model simulations.
- Comparison guidance: For offline land surface model comparisons, use RS+METEO products with corresponding meteorological forcing to minimize deviations from differing climate inputs.
- Setup limitations: RS products may be preferable for energy and water budget studies because they are not subject to uncertain meteorological inputs.Full consistency with forcing-specific simulations remains unattainable because FLUXCOM prescribes some seasonal and spatial land-surface properties, whereas the model simulates them.
- Temporal variability: Interannual variations are more uncertain than mean annual or seasonal spatial patterns, and monthly or annual anomalies should be normalized when comparing with simulations.FLUXCOM energy-flux interannual magnitudes are also likely too small.
- Budget accounting: Multiply FLUXCOM flux densities by the provided land fraction when calculating global or continental land budgets.Flux densities are defined per vegetated area, especially affecting sensible heat and net radiation accounting over non-vegetated areas.
- Uncertainty estimation: Multiply ensemble median absolute deviations by 1.4826 to obtain a robust normal-distribution standard-deviation estimate.Aggregating uncertainty across time or space requires assumptions about error covariances.
Data Citations
This section cites the Data Integration and Analysis System and the FLUXCOM EnergyFluxes dataset as related resources.
- Data Citations: Data Integration and Analysis System (DIAS) is cited with DOI 10.20783/DIAS.501 (2017).The citation credits Kim, H.
- Data Citations: FLUXCOM EnergyFluxes_v1 is cited with DOI 10.17871/FLUXCOM_EnergyFluxes_v1 (2018).The citation credits Jung, Koirala, Weber, Ichii, Gans, Camps-Valls, Papale, Schwalm, Tramontana, and Reichstein.