Quantifying high-resolution carbon flux in terrestrial ecosystems using geostationary satellites and knowledge-guided machine learning.

Quantitative estimation of terrestrial ecosystem carbon cycles is essential for advancing our understanding of climate change, given their role in absorbing approximately 30% of annual global anthropogenic CO2 emissions. To enhance the precision of carbon flux estimation, we developed a knowledge-guided machine learning model named the Carbon Simulator from Space (CASS), built upon the Light Use Efficiency (LUE) model. Leveraging primarily FLUXNET2015 data, our model trained and tested on parameters such as air temperature, relative humidity, photosynthetically active radiation (PAR), Enhanced Vegetation Index (EVI), and Land Surface Water Index (LSWI), and forest age to estimate LUE. For improved temporal and spatial resolution in carbon flux estimation, we incorporated refined datasets including PAR from the HIMAWARI8 geostationary satellite and reanalysis weather dataset, the Local Data Assimilation and Prediction System (LDAPS). The resulting output from CASS provides real-time estimates of carbon flux, offering hourly data with a spatial resolution of 250m, enabling the detection of land use changes associated with urbanization. Notably, CASS operates independently of Plant Function Type (PFT) dependency by replacing empirical coefficients for LUE with a machine learning regressor. These refined carbon flux estimates are anticipated to inform practical forest management strategies amidst the ongoing climate crisis.