Plants and ecosystems absorb and release carbon dioxide (CO₂) through processes like photosynthesis and respiration. These exchanges, called carbon fluxes, play a key role in climate change. Scientists measure carbon fluxes using ground stations, but these stations are spread out and don’t cover the whole globe, making it hard to get a complete picture.

In this study, we introduce DeepCarbon — a new global dataset that uses advanced machine learning models to estimate carbon fluxes everywhere on Earth. These models combine satellite and weather data and are trained using a mix of physical knowledge and deep learning techniques to improve accuracy. When tested on more than 100 locations that were not used during training, DeepCarbon gave better results than existing datasets, especially in areas with little data like tropical forests and the Arctic. The dataset provides daily carbon flux estimates at fine spatial resolution from 2001 to 2024.