Wei et al. (2026) A framework for long-term vegetation latent heat estimation and forecasting combining ERA5-land and Landsat data

Identification

• Journal: Agricultural and Forest Meteorology
• Year: 2026
• Date: 2026-03-02
• Authors: Yizhao Wei, Jinhui Huang
• DOI: 10.1016/j.agrformet.2026.111058

Research Groups

• College of Environmental Science and Engineering, Sino-Canada R&D Centre on Water and Environmental Safety, Nankai University
• Shenzhen Research Institute of Nankai University

Short Summary

This study developed a globally applicable framework integrating ERA5-Land reanalysis and Landsat data with machine learning to estimate and forecast monthly vegetation latent heat (LE) at 30 m resolution from 1984 to the present. It found Random Forest performed best for estimation and proposed two forecasting frameworks, LE-ML and LE-Direct, with varying performance based on training data availability.

Objective

• To develop a globally applicable framework for long-term, high-resolution vegetation latent heat (LE) estimation and forecasting by integrating ERA5-Land reanalysis and Landsat observations.

Study Configuration

• Spatial Scale: Global, with monthly LE estimates at 30 m resolution.
• Temporal Scale: From 1984 to the present (monthly estimates).

Methodology and Data

• Models used: Random Forest (for LE estimation), time series models (for LE-ML driver prediction), LE-ML framework, LE-Direct framework.
• Data sources: ERA5-Land reanalysis, Landsat observations (Normalized Difference Vegetation Index - NDVI), Fluxnet2015 tower data (for validation).

Main Results

• Six ERA5-Land variables and Landsat NDVI were identified as optimal drivers for LE estimation.
• Random Forest achieved the best performance for LE estimation (R² = 0.74; RMSE = 16.74 W m⁻²).
• Model skill for LE estimation declined with increasing NDVI extraction radius, reflecting spatial noise.
• For forecasting, LE-ML performed better with limited training data (mean R² = 0.62) compared to LE-Direct (mean R² = 0.48).
• LE-Direct outperformed LE-ML when trained on longer records (> 9 years) by effectively capturing seasonal cycles.

Contributions

• Development of a scalable and globally applicable framework for long-term (1984-present), high-resolution (30 m) vegetation LE estimation and forecasting.
• Integration of ERA5-Land reanalysis and Landsat data with machine learning for improved LE mapping.
• Introduction of two complementary forecasting frameworks (LE-ML and LE-Direct) with guidance on strategy selection based on data availability.

Funding

• Not specified in the provided text.

Citation

@article{Wei2026framework,
  author = {Wei, Yizhao and Huang, Jinhui},
  title = {A framework for long-term vegetation latent heat estimation and forecasting combining ERA5-land and Landsat data},
  journal = {Agricultural and Forest Meteorology},
  year = {2026},
  doi = {10.1016/j.agrformet.2026.111058},
  url = {https://doi.org/10.1016/j.agrformet.2026.111058}
}