Inferring global terrestrial carbon fluxes from the synergy of Sentinel 3 & 5P with Gaussian process hybrid models

• Published in: Remote sensing of environment Vol. 305; p. 114072
• Main Authors: Reyes-Muñoz, Pablo, D.Kovács, Dávid, Berger, Katja, Pipia, Luca, Belda, Santiago, Rivera-Caicedo, Juan Pablo, Verrelst, Jochem
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.05.2024
• Subjects: Gaussian process regression; Google Earth Engine; Hybrid models; OLCI; Photosynthesis; Sentinel; Solar-induced fluorescence; Terrestrial carbon fluxes; TROPOMI; Sentinel; Hybrid models; Google Earth Engine; OLCI; Solar-induced fluorescence; Gaussian process regression; Terrestrial carbon fluxes; TROPOMI; Photosynthesis
• ISSN: 0034-4257; 1879-0704
• DOI: 10.1016/j.rse.2024.114072

The ongoing monitoring of terrestrial carbon fluxes (TCF) goes hand in hand with progress in technical capacities, such as the next-generation Earth observation missions of the Copernicus initiative and advanced machine learning algorithms. Proceeding along this line, we present a physically-based data-driven workflow for quantifying gross primary productivity (GPP) and net primary productivity (NPP) at a global scale from the synergy of Copernicus’ Sentinel-3 (S3) Ocean and Land Color Instrument (OLCI) and the TROPOspheric Monitoring Instrument (TROPOMI) onboard Sentinel-5 Precursor (S5P), along with meteorological variables from Copernicus ERA5-Land. Specifically, we created generic hybrid Gaussian process regression (GPR) retrieval models combining S3-OLCI-derived vegetation products with the TROPOMI solar-induced fluorescence (SIF) product to capture global GPP and NPP. First, the GPR algorithms were trained on theoretical simulations through the Soil-Canopy-Observation of Photosynthesis and Energy fluxes (SCOPE) model, with the final retrieval models termed SCOPE-GPR-TCF. Second, the SCOPE-GPR-TCF models were integrated in Google Earth Engine (GEE) and fed with satellite data and products (coming from Sentinel 3 & 5P and ERA5-Land), producing global and regional (Iberian Peninsula) maps at spatial resolutions of 5 km and 300 m during the year 2019. Moderate relative uncertainties in the range between 10%–40% of the GPP and NPP estimates were achieved by the SCOPE-GPR-TCF models. Analysis of the driving variables revealed that the S3-OLCI vegetation products, i.e., leaf area index (LAI), the fraction of absorbed photosynthetically active radiation (FAPAR), and SIF provided the highest prediction strengths. Validation of GPP temporal estimates from GPR against partitioned GPP estimates at 113 flux towers located in America and Europe highlighted a good overall consistency at the local scale, with performances varying depending on the site and vegetation type. The highest scores emerged for stations located in croplands, grasslands, deciduous broad-leaf and evergreen needle-leaf forests with top R2 and rmse values above 0.8 and below 2 μmolm−2s−1 respectively. Further, benchmarking spatiotemporal analysis revealed a strong intra-annual global correlation against reference products for the same year 2019: (i) Cross-comparison against LPJ-GUESS resulted in modal values of R = 0.8 and rmse = 1.93 μmolm−2s−1 for GPP. (ii) MOD17A2H GPP and NPP estimations cross-correlated with R modal values of 0.94 and 0.92 and rmse of 1.26 and 1.05 μmolm−2s−1, respectively. We conclude that the hybrid models integrated into the GEE cloud-computing platform facilitate streamlining the global mapping of TCF products at efficient processing costs. This is particularly promising in preparation for the upcoming Fluorescence Explorer (FLEX) mission, where the SCOPE-GPR-TCF models are foreseen to be customized to 300 m resolution FLEX SIF data streams for high-resolution global productivity monitoring. [Display omitted] Workflow for modelling GPP and NPP based on a hybrid strategy: From top-bottom (1) Processing and standardizing data from diverse sources for feeding hybrid models. (2) SCOPE-GPR-TCF hybrid models, trained through SCOPE simulations and used for spatiotemporal mapping of wide areas through Google Earth Engine. (3) Validation with flux towers observations, LPJ-GUESS, and MOD17A2H products. •Synergy of Sentinel 3 & 5P for quantifying global terrestrial carbon fluxes.•Gaussian process regression for prediction and analysis of uncertainties.•Benchmarking against MODIS-based retrievals and process-based models.•SIF, LAI and FAPAR, retrievable from satellite, provided highest prediction strengths.•Results show the potential of Sentinel-3 & FLEX tandem for improving carbon estimates.