Wu et al. (2025) Multi-spatial scale assessment and multi-dataset fusion of global terrestrial evapotranspiration datasets

Identification

• Journal: Earth system science data
• Year: 2025
• Date: 2025-11-25
• Authors: Yi Wu, Chiyuan Miao, Yiying Wang, Qi Zhang, Jiachen Ji, Yuanfang Chai
• DOI: 10.5194/essd-17-6445-2025

Research Groups

State Key Laboratory of Earth Surface Processes and Disaster Risk Reduction, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China

Short Summary

This study comprehensively evaluates 30 global terrestrial evapotranspiration (ET) datasets across multiple spatial scales and fuses them using a Bayesian model averaging (BMA) method to create a new, robust, and long-term ET dataset (BMA-ET) for 1980–2020, demonstrating improved accuracy and capturing a global increasing ET trend.

Objective

• To comprehensively evaluate the accuracy and uncertainty of thirty existing global terrestrial evapotranspiration (ET) datasets (remote sensing-based, machine learning-based, reanalysis-based, and land surface model-based) across multiple spatial scales.
• To develop a new, more robust, and long-term global ET fusion dataset (BMA-ET) using a Bayesian model averaging (BMA) algorithm with a dynamic weighting scheme, addressing limitations of prior ET fusion efforts.

Study Configuration

• Spatial Scale: Global (domain of -89.75–89.75° N, -180–179.5° E), basin (32 basins), and site (174 FLUXNET2015 sites). Input datasets range from 0.05° × 0.05° to 1.875° × 1.875° resolution, standardized to 0.5° × 0.5° and 1° × 1° for evaluation and output BMA-ET.
• Temporal Scale: BMA-ET dataset covers 1980–2020. Common period for ET dataset evaluation is 1982–2011. Site-scale evaluation uses 1991–2011. Data has monthly resolution.

Methodology and Data

• Models used:
  • Fusion Method: Bayesian model averaging (BMA) with a dynamic weighting scheme.
  • Individual ET datasets (30 total, four types):
    • Remote Sensing-based: Penman-Monteith-Leuning (PML), Global Land Evaporation Amsterdam Model (GLEAM), Global Land Surface Satellite (GLASS), Process-based Land Surface Evapotranspiration/Heat Fluxes (PLSH).
    • Machine Learning-based: FLUXCOM (CRUNCEP_v8, GSWP3, WFDEI), Model Tree Ensemble Evapotranspiration (MTE).
    • Reanalysis-based: ERA5-Land, MERRA-Land.
    • Land Surface Model-based: GLDAS (CLSM, NOAH, VIC), TRENDY v12 models (CABLE-POP, CLASSIC, CLM5.0, DLEM, E3SM, EDv3, IBIS, ISBA-CTRIP, JSBACH, LPJ-GUESS, LPJmL, LPX-Bern, OCN, ORCHIDEE, SDGVM, VISIT, YIBs).
• Data sources:
  • Reference ET Observations: FLUXNET2015 (212 flux sites, 1991–2014).
  • Independent Validation Data: AmeriFlux (1994–2020), ChinaFlux (2003–2010), ICOS (2003–2010).
  • Basin Data: HydroBasins Level 3 basin boundaries, Dai and Trenberth Global River Flow and Continental Discharge Dataset, Global Runoff Data Centre (streamflow).
  • Precipitation Data: Global Precipitation Climatology Centre (GPCC, 1° × 1°).
  • Land Cover Data: MODIS MCD12Q1 (500 m resolution, IGBP classification).

Main Results

• The BMA-ET fusion dataset was generated for 1980–2020 at 0.5° × 0.5° and 1° × 1° spatial resolutions.
• Global terrestrial BMA-ET shows an overall increasing trend of 0.65 mm yr⁻¹ (with a 95% confidence interval of 0.51–0.78 mm yr⁻¹) during 1980–2020 (p < 0.01).
• BMA-ET exhibits higher correlation coefficients and lower root-mean-square errors compared to most individual ET datasets.
• Validation using FLUXNET2015 as reference shows that over 70% of flux sites have correlation coefficients exceeding 0.6 with BMA-ET.
• Validation against independent data sources demonstrates strong performance: correlation coefficients of BMA-ET with AmeriFlux, ChinaFlux, and ICOS reach 0.61, 0.72, and 0.74, respectively.
• Remote sensing- and machine learning-based ET datasets (e.g., MTE, PML, PLSH) generally show superior accuracy at the site scale, though optimal selection depends on season and vegetation type.
• At the basin scale, most ET datasets demonstrate superior performance, with R² values exceeding 0.8, and often 0.9, when compared to water balance-based observations.
• Machine learning- and remote sensing-based datasets generally exhibit lower relative uncertainty at the grid point scale.

Contributions

• This study presents the first ET fusion dataset that integrates all four main types of global ET datasets (remote sensing-based, machine learning-based, reanalysis-based, and land surface model-based), addressing the limitation of prior fusion efforts that typically relied on a single type.
• It introduces a novel dynamic weighting scheme within the Bayesian model averaging (BMA) algorithm, which adjusts weights based on different vegetation types and years with non-overlapping coverage among ET datasets, significantly improving the utilization of FLUXNET2015 observations.
• Provides a comprehensive, long-term (1980–2020) global terrestrial ET dataset (BMA-ET) with improved accuracy and reduced uncertainty, serving as a valuable resource for understanding global ET patterns and trends.
• Conducts a multi-spatial scale (site, basin, global) assessment of 30 diverse ET datasets, offering detailed insights into their performance across different conditions.

Funding

• National Key Research and Development Program of China (2024YFF0809301)
• National Natural Science Foundation of China (U24A20572)
• Fundamental Research Funds for the Central Universities

Citation

@article{Wu2025Multispatial,
  author = {Wu, Yi and Miao, Chiyuan and Wang, Yiying and Zhang, Qi and Ji, Jiachen and Chai, Yuanfang},
  title = {Multi-spatial scale assessment and multi-dataset fusion of global terrestrial evapotranspiration datasets},
  journal = {Earth system science data},
  year = {2025},
  doi = {10.5194/essd-17-6445-2025},
  url = {https://doi.org/10.5194/essd-17-6445-2025}
}