Enhancing Carbon Cycle Understanding through Deep Learning: Development and Validation of the Global Carbon Fluxes Dataset (GCFD)

This research tackles the constraints inherent in current global carbon flux datasets and introduces a groundbreaking new dataset, the Global Carbon Fluxes Dataset (GCFD), which integrates cutting-edge deep learning methodologies alongside in situ measurements. GCFD delivers unprecedented high-resolution spatial and temporal data on Gross Primary Productivity (GPP), Terrestrial Ecosystem Respiration (RECO), and Net Ecosystem Exchange (NEE). The Convolutional Neural Network (CNN) model employed in this study surpasses conventional machine learning techniques, demonstrating robust performance in modeling GPP, RECO, and NEE.

The precision and spatial granularity of GCFD outshine those of alternative global carbon flux datasets, like FLUXCOM, and it exhibits strong coherence with remote sensing vegetation condition data. Serving as a reliable reference for both meteorological and ecological investigations, GCFD is particularly valuable when high-resolution carbon flux mapping is essential. Its reliability has been rigorously tested by comparative analysis against existing data products, revealing insightful details about the global spatial and temporal patterns of carbon fluxes, especially within tropical and dry climate zones where notable trends have emerged.

This study significantly advances our comprehension of worldwide carbon flux dynamics and underscores the untapped potential of deep learning technologies to enhance the quality of carbon flux datasets. Accessible at https://dx.doi.org/DOI:10.11888/Terre.tpdc.300009, GCFD offers data resolutions ranging from 1 km to 9 km.