Liu et al. (2026) Multi-satellite data fusion for improved field-scale evapotranspiration mapping on Google Earth Engine

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-02-13
• Authors: Hui Liu, Yun Yang, Feng Gao, Feng Gao, Christopher R. Hain, Vikalp Mishra, John Volk, Yanghui Kang
• DOI: 10.1016/j.rse.2026.115299

Research Groups

• Soil and Crop Sciences Section, School of Integrative Plant Science, Cornell University, Ithaca, NY, USA
• USDA, Agricultural Research Service, Hydrology and Remote Sensing Laboratory, Beltsville, MD, USA
• NASA, Marshall Space Flight Center, Earth Science Branch, Huntsville, AL, USA
• Earth System Science Center, The University of Alabama in Huntsville, Huntsville, AL, USA
• NASA Short-term Prediction Research and Transition Center, Marshall Space Flight Center, Huntsville, AL, USA
• Desert Research Institute, Reno, NV, USA
• Department of Biological Systems Engineering, Virginia Tech, Blacksburg, VA, USA

Short Summary

This study developed a Google Earth Engine (GEE)-based framework to improve field-scale evapotranspiration (ET) mapping by fusing thermal infrared (TIR) observations from ECOSTRESS and VIIRS with Harmonized Landsat-Sentinel (HLS) data. The integration of multi-satellite data generally enhanced ET estimation accuracy, reducing average Mean Absolute Error (MAE) by 8.64% (daily), 14.40% (weekly), and 16.37% (monthly) compared to Landsat-only baselines.

Objective

• To evaluate the feasibility of combining Harmonized Landsat-Sentinel (HLS), ECOSTRESS, and VIIRS data for high-frequency, 30-meter resolution evapotranspiration (ET) mapping within the Google Earth Engine (GEE) environment for operational data production, using the OpenET GEE-DisALEXI implementation as a testbed.
• To integrate the 30-meter HLS product, the 70-meter ECOSTRESS Collection 2 Land Surface Temperature (LST) product, and the 375-meter VIIRS LST product into GEE-DisALEXI.
• To evaluate the performance of various data fusion combinations (Landsat-only, Landsat+ECOSTRESS, and Landsat+ECOSTRESS+VIIRS) by comparing them with flux tower observations.
• To analyze the operational challenges and limitations of using LST data from multiple satellites for ET retrieval on GEE.

Study Configuration

• Spatial Scale: Field-scale (30-meter resolution) over six AmeriFlux sites in California, Missouri, and North Carolina, representing diverse land cover and climatic conditions. Potential CONUS-wide application assessed.
• Temporal Scale: Daily, weekly, and monthly evapotranspiration (ET) estimates during the growing season (April to September) for the years 2020, 2022, and 2023. Satellite revisit intervals: Landsat (~8 days), ECOSTRESS (~4 days average), VIIRS (daily), HLS (3–4 days effective).

Methodology and Data

• Models used:
  • GEE-based Disaggregated Atmosphere-Land Exchange (DisALEXI) model (based on the Two-Source Energy Balance - TSEB).
  • Data Mining Sharpener (DMS) algorithm (modified for GEE and multi-sensor LST sharpening).
  • ALEXI model (for coarse-scale ET constraint).
  • Random Forest regression (within DMS for global LST-SR relationship).
  • Ordinary Least Squares (OLS) regression (within DMS for local LST-SR relationship).
  • American Society of Civil Engineers (ASCE) Standardized Penman-Monteith equation (for reference ET).
• Data sources:
  • Satellite:
    • Landsat Collection 2 (Landsat 8/9): 30-meter Visible/Shortwave Infrared (VSWIR) and 100-meter Thermal Infrared (TIR) Land Surface Temperature (LST).
    • Harmonized Landsat-Sentinel (HLS) dataset (HLSS30 from Sentinel-2 A/B, HLSL30 from Landsat 8/9): 30-meter surface reflectance (VSWIR).
    • ECOSTRESS (on ISS) Collection 2: ~70-meter LST and Emissivity (LST&E) product (LST layer).
    • Visible Infrared Imaging Radiometer Suite (VIIRS) (on Suomi-NPP and JPSS-1): 375-meter LST (from I5 band).
    • Geostationary Operational Environmental Satellites (GOES): 4-kilometer ALEXI daily ET data.
  • Meteorological:
    • North American Land Data Assimilation System (NLDAS) forcing dataset: 2-meter air temperature.
    • Climate Forecast System Reanalysis (CFSR): Hourly insolation, three-hourly wind speed, air temperature, atmospheric pressure, and vapor pressure.
    • gridMET dataset: Meteorological inputs for reference ET.
  • Auxiliary/Ground:
    • 2019 National Land Cover Dataset (NLCD): 30-meter land cover classification.
    • AmeriFlux network: Eddy covariance flux tower observations (daily ET, closure-corrected ET, average ET).
    • Phenocam imagery: Green Chromatic Coordinate (GCC) for canopy greenness.

Main Results

• The integration of ECOSTRESS and VIIRS data generally improved ET estimation accuracy across most sites and temporal scales compared to Landsat-only baselines.
• Average Mean Absolute Error (MAE) (in millimeters per day) was reduced by 8.64% (from 1.12 to 1.02) for daily estimates, 14.40% (from 1.00 to 0.85) for weekly estimates, and 16.37% (from 0.82 to 0.69) for monthly estimates.
• Multi-satellite data fusion effectively captured rapid ET declines following cutting events (e.g., alfalfa fields) or vegetation cover transitions, which were often missed by Landsat-only data due to insufficient temporal sampling.
• Spatial assessments demonstrated strong coherence and structural fidelity in ET retrievals across different satellite platforms after Data Mining Sharpener (DMS) processing, with differences generally within ±1 millimeter per day.
• ECOSTRESS ET retrievals exhibited a clear dependence on sensor view angle and overpass time, with accuracy degrading at large view angles (exceeding 20 degrees) and during early morning (before 09:00 local time) or late afternoon (after 17:00 local time) acquisitions.
• VIIRS ET retrievals showed a weaker dependence on view angle and no systematic dependence on overpass time, attributed to its consistent early-afternoon acquisition window.
• A CONUS-wide assessment indicated that integrating ECOSTRESS and VIIRS substantially increased the number of effective observation days, with most regions exceeding 60 days (compared to 10–50 days for Landsat alone), and some western US regions reaching over 100–180 days.

Contributions

• Developed and implemented a novel, scalable Google Earth Engine (GEE)-based framework for high-frequency, 30-meter resolution evapotranspiration (ET) mapping by fusing Harmonized Landsat-Sentinel (HLS), ECOSTRESS, and VIIRS data.
• Adapted the Data Mining Sharpener (DMS) algorithm for efficient multi-sensor Land Surface Temperature (LST) sharpening within the GEE environment, maintaining a dual-model structure to capture local heterogeneity while addressing GEE computational constraints.
• Provided a comprehensive evaluation of the benefits of multi-satellite data fusion for ET estimation across diverse land cover and climatic conditions, including cloud-prone humid eastern US sites, which extends previous studies focused primarily on irrigated areas.
• Established a prototype workflow for integrating new satellite data sources into the OpenET modeling framework, thereby supporting sustainable agriculture and water resource management.
• Quantified the impact of sensor-specific characteristics, such as view angle and acquisition time, on ET retrieval accuracy for ECOSTRESS and VIIRS within the GEE-DisALEXI framework.

Funding

• NASA ECOSTRESS program (Grant Number: 80NSSC23K0642)
• USGS OpenET Project Cooperative Agreement (G23AC00676)
• NASA ACRES project (Grant Number: 80NSSC23M0034)

Citation

@article{Liu2026Multisatellite,
  author = {Liu, Hui and Yang, Yun and Gao, Feng and Gao, Feng and Hain, Christopher R. and Mishra, Vikalp and Volk, John and Kang, Yanghui},
  title = {Multi-satellite data fusion for improved field-scale evapotranspiration mapping on Google Earth Engine},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115299},
  url = {https://doi.org/10.1016/j.rse.2026.115299}
}