Treffer: High-spatial-resolution gross primary production estimation from Sentinel-2 reflectance using hybrid Gaussian processes modeling.

High-spatial-resolution gross primary production (GPP) estimation is critical for local carbon monitoring, especially in heterogeneous landscapes where global products lack spatial detail. We present a hybrid modeling framework that estimates GPP using Sentinel-2 (S2) reflectance and Bayesian Gaussian Process Regression (GPR), chosen for its robustness with limited data and its ability to quantify uncertainty. GPR models were trained using SCOPE (Soil Canopy Observation of Photosynthesis and Energy fluxes) radiative transfer model (RTM) simulations and optimized via active learning (AL) across 10 plant functional types (PFTs). These lightweight, PFT-specific S2-GPR models were implemented in Google Earth Engine (GEE) to enable scalable, reproducible, and accessible GPP estimation and mapping. S2-GPR models predictive performances were evaluated using data from 67 eddy covariance flux towers across Europe. Data from 2017–2020 were used for training and training database optimization, while 2021–2024 data served as independent validation. Strong predictive performance was achieved 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), with lower accuracy in evergreen broadleaf (R=0.07) and needleleaf forests (R=0.33). Compared to MODIS GPP (MOD17A2H V6.1), the S2-GPR models showed consistently lower bias and comparable or improved accuracy in most PFTs, except evergreen forests. Additional validation against AmeriFlux sites in North America demonstrated that the models retain predictive power beyond the ICOS network, though ecosystem-specific and regional differences can influence accuracy. The inclusion of coarse-resolution meteorological variables (temperature, radiation, vapor pressure deficit, air pressure) was evaluated but generally did not improve predictive performance and introduced additional uncertainty, highlighting that in this study S2 spectral information alone provides the dominant signal for high-resolution GPP estimation. These findings underscore the value of integrating SCOPE modeling and AL-optimized GPR for accurate, local-scale GPP mapping using cloud-based S2 data, complementing coarse-resolution global products. [Display omitted] • High-spatial-resolution GPP estimation from Sentinel-2 imagery in Google Earth Engine. • Hybrid Gaussian Processes modeling with SCOPE and optimized via active learning. • Development of ten Plant Functional Type-specific S2-GPR models for GPP estimation. • Trained and tested on 67 European ICOS flux towers and additionally validated with AmeriFlux sites. • Comparative performance assessment against the benchmark MODIS GPP product (MOD17A2H). • Inclusion of ERA5-Land meteorological variables resulted in minimal improvement. [ABSTRACT FROM AUTHOR]