Clerck et al. (2025) High-spatial-resolution gross primary production estimation from Sentinel-2 reflectance using hybrid Gaussian processes modeling

Identification

• Journal: ISPRS Journal of Photogrammetry and Remote Sensing
• Year: 2025
• Date: 2025-12-10
• Authors: Emma De Clerck, Pablo Reyes-Muñoz, Egor Prikaziuk, Dávid D.Kovács, Jochem Verrelst
• DOI: 10.1016/j.isprsjprs.2025.11.033

Research Groups

• Image Processing Laboratory (IPL) - University of Valencia, Paterna, Spain
• Faculty of Geo-Information Science and Earth Observation (ITC) - University of Twente, Enschede, The Netherlands
• Department of Geodesy and Geoinformation, TU Wien, Vienna, Austria

Short Summary

This study develops a hybrid modeling framework using Sentinel-2 reflectance and Gaussian Process Regression (GPR) trained with SCOPE radiative transfer model simulations to estimate high-spatial-resolution (20 meters) Gross Primary Production (GPP) across 10 plant functional types (PFTs). The PFT-specific GPR models, implemented in Google Earth Engine, demonstrated strong predictive performance in most ecosystems, outperforming MODIS GPP in terms of bias and spatial detail, though showing limitations in evergreen forests.

Objective

• Build 20 meter resolution Sentinel-2 Gaussian Process Regression (S2-GPR) models optimized by Plant Functional Type (PFT).
• Test and validate these S2-GPR models using eddy covariance (EC) flux tower observations and MODIS Gross Primary Production (GPP) products.
• Enable accessible, fine-scale GPP mapping across diverse landscapes through an open-source Google Earth Engine (GEE) workflow.

Study Configuration

• Spatial Scale: Sentinel-2 data at 10 and 20 meters resolution (resampled to 20 meters); MODIS GPP at 500 meters; ERA5-Land reanalysis at approximately 11 kilometers; EC flux tower buffers of 100 and 200 meters radius; GPP mapping demonstrations over 1 kilometer radius.
• Temporal Scale: Sentinel-2 revisit time of approximately 5 days; MODIS GPP 8-day composites; ICOS and AmeriFlux EC flux tower data aggregated to daily GPP; Data period from 2017–2024 (2017–2020 for training/testing, 2021–2024 for independent validation).

Methodology and Data

• Models used:
  • Hybrid modeling framework combining Gaussian Process Regression (GPR) with SCOPE (Soil Canopy Observation of Photosynthesis and Energy fluxes) radiative transfer model (RTM) simulations (v1.90).
  • Active Learning (AL) for optimizing GPR training datasets.
  • Google Earth Engine (GEE) for scalable implementation and processing.
  • ARTMO (Automatic Radiative Transfer Model Operator) Matlab software package for model building.
  • PyEOGPR Python package for GPR model deployment.
• Data sources:
  • Satellite: Sentinel-2 (S2) Level 2A (L2A) surface reflectance (10 spectral bands at 10 and 20 meters resolution, resampled to 20 meters).
  • Observation:
    • Integrated Carbon Observation System (ICOS) eddy covariance (EC) flux tower network (67 sites across Europe) for daily GPP (2017–2024).
    • AmeriFlux network (North America) for independent validation (2017–2024).
  • Reanalysis: ERA5-Land reanalysis data (approximately 11 kilometers resolution) for meteorological variables (daily temperature in degrees Celsius, incoming shortwave radiation in watts per square meter, vapor pressure deficit in hectopascals, air pressure in hectopascals).
  • Benchmark Product: MODIS GPP (MOD17A2H Version 6.1) at 500 meters spatial resolution, 8-day temporal composite.

Main Results

• PFT-Specific Performance (2017–2020 testing): S2-GPR models achieved strong predictive performance in wetlands (R=0.84, NRMSE=12.6%), savannas (R=0.81, NRMSE=12.2%), and deciduous broadleaf forests (R=0.81, NRMSE=14.3%). Moderate accuracy was observed for croplands, shrublands, grasslands, and mixed forests (R=0.67–0.77, NRMSE=14%–18%). Lower accuracy was found in evergreen broadleaf (R=0.07, NRMSE=25.6%) and needleleaf forests (R=0.33, NRMSE=19.6%), where models struggled to capture physiological variability.
• Independent Validation (2021–2024) vs. MODIS GPP: S2-GPR models consistently showed lower bias and comparable or improved accuracy in most PFTs (e.g., croplands, deciduous broadleaf forests, mixed forests) compared to MODIS GPP. However, MODIS GPP demonstrated superior accuracies in evergreen broadleaf and needleleaf forests.
• Spectral Band Importance: Analysis revealed that blue (B2) and red (B4) bands were consistently important for most PFTs. Red-edge bands (B5, B6, B7, B8A) showed differential importance, with B5 and B6 being particularly important for deciduous broadleaf forests. Near-infrared (B8) was important for grasslands and wetlands, and Shortwave Infrared (SWIR) bands (B11, B12) were consistently important across all PFTs.
• Impact of Meteorological Variables: The inclusion of coarse-resolution meteorological variables from ERA5-Land generally did not improve S2-GPR model performance and often introduced additional uncertainty, suggesting that S2 spectral information alone provides the dominant signal for high-resolution GPP estimation in this context.
• High-Spatial-Resolution Mapping: S2-GPR models successfully captured fine-scale spatial heterogeneity and temporal variability in GPP at 20 meters resolution, outperforming MODIS GPP (500 meters) in heterogeneous and mixed-pixel landscapes.
• AmeriFlux Validation: Independent validation against AmeriFlux sites in North America confirmed reasonable generalization for deciduous broadleaf forests and grasslands, but highlighted persistent challenges in evergreen forests, wetlands, and shrublands, indicating sensitivity to ecosystem-specific and regional differences.

Contributions

• Development of Plant Functional Type (PFT)-specific and scalable hybrid Gaussian Process Regression (GPR) models for Sentinel-2 (S2) reflectance data, requiring no additional meteorological inputs.
• Creation of lightweight, PFT-optimized models designed for efficient execution in cloud computing platforms like Google Earth Engine (GEE), avoiding local data download and processing.
• Provision of explicit uncertainty intervals for Gross Primary Production (GPP) estimates, a crucial feature often absent in operational products.
• Demonstration of the value of integrating SCOPE radiative transfer modeling and Active Learning (AL)-optimized GPR for accurate, local-scale GPP mapping using cloud-based S2 data, complementing coarse-resolution global products.
• Comprehensive evaluation of the full Sentinel-2 spectral range, highlighting the importance of red-edge and Shortwave Infrared (SWIR) bands for GPP estimation across diverse PFTs.

Funding

• European Research Council (ERC) under the FLEXINEL project (grant number 101086622).
• EU COST (European Cooperation in Science and Technology) Action CA22136 ‘‘Pan-European Network of Green Deal Agriculture and Forestry Earth Observation Science (PANGEOS)’’.

Citation

@article{Clerck2025Highspatialresolution,
  author = {Clerck, Emma De and Reyes-Muñoz, Pablo and Prikaziuk, Egor and D.Kovács, Dávid and Verrelst, Jochem},
  title = {High-spatial-resolution gross primary production estimation from Sentinel-2 reflectance using hybrid Gaussian processes modeling},
  journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
  year = {2025},
  doi = {10.1016/j.isprsjprs.2025.11.033},
  url = {https://doi.org/10.1016/j.isprsjprs.2025.11.033}
}