Machine Learning Produces Unprecedented High-Resolution Map of Global Soil Respiration
Scientists at the U.S. Department of Energy’s Pacific Northwest National Laboratory have developed and continue to maintain a global database of measurements made of soil-to-atmosphere CO2 flows, termed soil respiration. A research team at the University of Delaware has leveraged these observations in a machine-learning approach to create a new high-resolution global map of soil respiration and its uncertainties.
Soil respiration is one of the largest fluxes in the global carbon cycle, providing critical insights into biological activity in the underlying soil. This new global map of soil respiration and its uncertainties provides modelers and experimentalists with a “gold standard” benchmark dataset identifying areas with the highest uncertainties to target in the future.
Soils emit large amounts of carbon dioxide into the atmosphere every year via the process of soil respiration. Rates of soil respiration are highly variable in space, however, limiting scientists’ ability to balance global carbon budgets and forecast climate change. This study used a novel machine learning approach to predict soil respiration rates at high resolution (1 km2) globally, based on how observations of soil respiration were related to climate (annual temperature, annual and seasonal precipitation) and vegetation. It also examined the spatial patterns of the associated uncertainty of these predictions. Predicted annual soil respiration and prediction uncertainty varied across ecosystem types and regions, with evergreen tropical forests dominating global annual soil respiration emissions. Dryland, wetland, and cold ecosystems had the highest associated prediction uncertainties, suggesting that future soil respiration measurements would be especially useful in these areas. The high spatial resolution of these predictions will help researchers studying the carbon cycle at local to global scales and provide a high-quality benchmark dataset for Earth System Models.