Wang et al. (2025) GPP-net: a robust high-resolution GPP estimation network for Sentinel-2 using only surface reflectance and photosynthetically active radiation

Identification

• Journal: Remote Sensing of Environment
• Year: 2025
• Date: 2025-12-16
• Authors: Shaoyu Wang, Youngryel Ryu, Benjamin Dechant, Helin Zhang, Huaize Feng, Jeongho Lee, Changhyun Choi
• DOI: 10.1016/j.rse.2025.115198

Research Groups

• Research Institute of Agriculture and Life Sciences, Seoul National University, Seoul, Republic of Korea
• Interdisciplinary Program in Agricultural and Forest Meteorology, Seoul National University, Seoul, Republic of Korea
• Department of Landscape Architecture and Rural Systems Engineering, Seoul National University, Seoul, Republic of Korea
• SNU Energy Initiatives, Seoul National University, Seoul, Republic of Korea
• German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig, Leipzig, Germany
• Leipzig University, Leipzig, Germany

Short Summary

This study introduces GPP-net, a deep learning network for robust, high-resolution gross primary productivity (GPP) estimation using only Sentinel-2 surface reflectance and photosynthetically active radiation (PAR). GPP-net demonstrates superior accuracy and generalization across diverse vegetation types and extreme climate conditions, significantly reducing reliance on traditional land cover and coarse meteorological data.

Objective

• Can we achieve robust, high-resolution GPP estimation without requiring land cover and other meteorological data?
• Can GPP-net achieve robust GPP estimation across plant functional types (PFTs) and C3 and C4 vegetations?
• Can GPP-net estimate GPP during droughts and heatwaves?

Study Configuration

• Spatial Scale: High-resolution (10-30 meters), utilizing Sentinel-2 Multispectral Instrument (MSI) data (10 m, 20 m, 60 m bands). GPP mapping was conducted at the pixel scale, with flux tower observations matched using 80% accumulated flux footprints or 3x3 pixel averages.
• Temporal Scale: Half-hourly and daily GPP estimation. Data collected from 2018 to 2023 (ICOS), 2018 to 2021 (AmeriFlux), and the FLUXNET2015 dataset. The study also evaluated inter-annual variability.

Methodology and Data

• Models used:
  • GPP-net: A novel deep learning network based on a fully 1-D convolutional encoder-decoder architecture combined with a spectral band importance estimation module (Squeeze-and-Excitation block).
  • SCOPE model: Soil-Canopy Energy Balance Radiative Transfer model, used to simulate reflectance data for pre-training GPP-net.
  • Benchmark models: Near-infrared reflectance of vegetation multiplied by incoming sunlight (NIRvP)-based linear regression, Partial Least Squares (PLS) regression, and Random Forest (RF) regression. These were tested with and without Plant Functional Type (PFT) information.
• Data sources:
  • Satellite data: Sentinel-2 Multispectral Instrument (MSI) surface reflectance (COPERNICUS/S2SRHARMONIZED product, bands B1, B2, B3, B4, B5, B6, B7, B8, B8A). Moderate Resolution Imaging Spectroradiometer (MODIS) Leaf Area Index (LAI) product (MODIS/061/MCD15A3H) for SCOPE simulations.
  • In-situ data: Gross primary productivity (GPP) and photosynthetically active radiation (PAR) from ICOS station network (62 flux towers), AmeriFlux network (29 sites), and FLUXNET2015 dataset (102 sites). Warm Winter 2020 dataset was used for evaluating performance during droughts and heatwaves. Meteorological data (e.g., vapor pressure deficit, air temperature, wind speed) from FLUXNET2015 were used as inputs for the SCOPE model.

Main Results

• GPP-netSCOPE (GPP-net with hybrid-SCOPE pre-training) achieved the highest half-hourly GPP estimation accuracy for Croplands, Wetlands, Evergreen Needleleaf Forest, Mixed Forest, and Deciduous Broadleaf Forest, and comparable accuracy for Grasslands and Evergreen Broadleaf Forest, outperforming benchmark models.
• GPP-netSCOPE demonstrated the highest overall half-hourly GPP estimation accuracy across all PFTs (R² increased by 0.04, Root Mean Square Error (RMSE) reduced by 0.6 µmol CO₂⋅m⁻²⋅s⁻¹, Relative RMSE (RRMSE) reduced by 4.42 percentage points compared to GPP-net without pre-training).
• GPP-netSCOPE consistently achieved the highest daily GPP estimation accuracy for Croplands and Wetlands, and the highest overall accuracy across all PFTs (R² increased by 0.06, RMSE reduced by 0.31 gC⋅m⁻²⋅d⁻¹, RRMSE reduced by 5.45 percentage points compared to GPP-net without pre-training).
• The red edge spectral band (B7) was identified as the most important for GPP estimation, followed by the Near-Infrared (B8) and Aerosols (B1) bands, with pre-training stabilizing this importance ranking.
• GPP-netSCOPE showed improved robustness to soil effects in GPP mapping and generated more reliable spatial patterns compared to other methods.
• GPP-net demonstrated a robust ability to capture inter-annual variability of GPP, achieving lower mean RRMSE and Maximum Absolute Deviation (MAD) values than benchmark linear and Random Forest methods.
• GPP-netSCOPE achieved robust GPP estimation for both C3 and C4 vegetation, exhibiting the smallest slope deviation between in-situ and estimated GPP and the highest prediction accuracy, unlike benchmark models that showed significant differences between C3 and C4.
• GPP-netSCOPE obtained the highest accuracy for GPP estimation under drought and heatwave events (R² = 0.46, RRMSE = 50.26%), with only marginal improvement from including Vapor Pressure Deficit (VPD) as a predictor, indicating its implicit capture of stress conditions.

Contributions

• Proposes GPP-net, a novel deep learning model that enables robust, high-resolution GPP estimation using only Sentinel-2 surface reflectance and photosynthetically active radiation, significantly reducing the need for land cover and coarse meteorological data.
• Introduces a unique pre-training strategy utilizing a hybrid-SCOPE dataset (SCOPE-simulated reflectance combined with FLUXNET2015 GPP and PAR) to enhance the generalization and robustness of the deep learning model.
• Demonstrates superior and consistent GPP estimation accuracy across seven diverse plant functional types and both C3 and C4 vegetation using a single, unified model, addressing a long-standing challenge in GPP modeling.
• Shows that GPP-net can reliably estimate GPP under extreme climate conditions (droughts and heatwaves) without explicit meteorological inputs like VPD, leveraging its advanced feature extraction capabilities.
• Confirms the model's ability to generate GPP maps that are robust against soil background effects and accurately track inter-annual variability of GPP.
• Provides a flexible deep learning framework that can continuously improve its generalization as new observational data become available, paving the way for future global high-resolution GPP mapping efforts.

Funding

• Ministry of Environment of Korea (2022003640002)
• Rural Development Administration of Korea (RS-2024-00397146)
• sDiv, the Synthesis Centre of the German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig (DFG FZT 118, 202548816)

Citation

@article{Wang2025GPPnet,
  author = {Wang, Shaoyu and Ryu, Youngryel and Dechant, Benjamin and Zhang, Helin and Feng, Huaize and Lee, Jeongho and Choi, Changhyun},
  title = {GPP-net: a robust high-resolution GPP estimation network for Sentinel-2 using only surface reflectance and photosynthetically active radiation},
  journal = {Remote Sensing of Environment},
  year = {2025},
  doi = {10.1016/j.rse.2025.115198},
  url = {https://doi.org/10.1016/j.rse.2025.115198}
}