Rapid changes in terrestrial carbon dioxide uptake captured in near-real time from a geostationary satellite: The ALIVE framework

The terrestrial carbon cycle responds to human activity, ecosystem dynamics, and weather and climate variability including extreme events. Satellite remote sensing has transformed our ability to estimate ecosystem carbon dioxide uptake, the gross primary productivity (GPP), with increasing accuracy and spatial resolution. Many aspects of terrestrial carbon cycling happen quickly on sub-daily or daily scales. These dynamics may not be captured at the temporal scales of typical remote sensing products from polar orbiting satellites – often multiple days or longer. Imagers onboard geostationary satellites measure the Earth system at “hypertemporal” time scales of minutes or less and often have the spectral capabilities to estimate GPP and other surface-atmosphere fluxes using established approaches. Here, we use observations and data products from the Advanced Baseline Imager (ABI) on the Geostationary Environmental Operational Satellite – R Series (GOES-R) to create ALIVE GPP ( A dvanced Baseline Imager L ive I maging of V egetated E cosystems), a GPP product that provides open data on the native five-minute basis of GOES-R CONUS scenes with latency under one day. Our machine learning model, trained on GPP estimates from 111 eddy covariance flux towers with 276 site-years of data spanning tropical to boreal ecosystems, captures up to 70 % of the observed variability when 20 % of tower sites are withheld, with R 2 values of 0.78 (0.82) when aggregating to daily (weekly) periods. We compared ALIVE GPP predictions against eight-day MODIS MOD17A2 GPP estimates and daily GPP estimates from the Breathing Earth System Simulator v2 (BESSv2) and demonstrate how ALIVE GPP simulates the impacts of phenological transitions, flash drought, and hurricanes. Advancements to geostationary satellite imagery, machine learning, and cloud computing make it possible to estimate carbon flux in near real-time and provide new ways to understand the ever-changing carbon cycle and the processes that define it. • We created ‘ALIVE GPP ’, a five-minute GPP mapping framework with <1 day latency. • Gradient boosting models were trained on eddy covariance and GOES-R data across CONUS. • ALIVE GPP R 2 is 0.67 to 0.70 for half-hourly intervals and 0.78 for daily sums. • Rapid regional GPP responses to extreme events matched expectations.