ULSAN — Researchers have developed an artificial intelligence technology that combines different satellite observations to estimate daily how much carbon dioxide plants absorb through photosynthesis, down to 30-meter units.
Ulsan National Institute of Science and Technology said on the 20th that a team led by Professor Im Jung-ho of the Department of Civil, Urban, Earth and Environmental Engineering has developed UNIQUE, an AI technology that analyzes satellite and weather data to produce maps of plant carbon uptake, measured as gross primary production (GPP).
Ground-based observation towers, known as flux towers, that measure plant carbon uptake cover only a narrow area. Satellite observation, used to fill that gap, has also struggled to deliver both the spatial resolution needed to examine small areas in detail and the temporal resolution needed to observe the same area frequently. The difficulty has been especially acute over Korean terrain, where forests and farmland are interspersed and summer cloud cover is heavy, making continuous high-precision data hard to obtain.
UNIQUE consists of two satellite-specific models for estimating carbon uptake and a generative AI-based fusion model, U-SCALER, that merges the two resulting maps into one.
The team first built dedicated estimation models for Landsat, which observes the land surface in detail at 30-meter units, and MODIS, which provides daily observations, drawing on data from some 300 observation sites across the Northern Hemisphere. The models combine NIRvP — an index multiplying satellite-measured near-infrared reflectance of vegetation by the amount of light used in photosynthesis — with temperature and moisture conditions.
U-SCALER then merges the carbon uptake maps produced by the two satellites to complete a daily carbon uptake map on a 30-by-30-meter grid. Areas left blank by cloud cover or other obstructions are corrected using data from preceding and following dates together with learned relationships, allowing maps to be generated without gaps even on days when Landsat makes no observations.
In performance testing, the technology distinguished seasonal differences in carbon uptake between a deciduous broadleaf forest and an evergreen coniferous forest about 320 meters apart at Mount Taehwa in Korea. Existing data at 500-meter units showed the two forests absorbing carbon at nearly identical rates, but the new technology reproduced differences by tree species in a way that closely matched actual measurements. Validation in Wisconsin in the United States also broke down carbon uptake differences in detail across an area where farmland and forest were mixed.
Estimation accuracy also improved. Daily errors from the satellite-specific estimation models were 13% to 28% lower than those of existing global carbon uptake datasets. When the final merged map was compared with measurements from 30 ground observation stations, the root mean square error came to 2.60 grams of carbon per square meter per day, about 8% lower than that of the existing image fusion technique known as ESTARFM.
The methodology has been optimized and is now being applied to the land carbon uptake diagnosis section of the Korea Carbon Project, a region-specific carbon information platform.
"If we expand the scope by applying a wider range of satellites and carbon-related variables, this can be used for precision monitoring to support forest and farmland management, assessments of ecosystem impacts from climate disasters, and the formulation of carbon neutrality policy," Im said.
Son Bo-kyung and Sung Tae-jun, researchers at UNIST, served as co-first authors, with participating researchers from Seoul National University, Pusan National University and the University of Freiburg in Germany. The findings were published online in the international journal Remote Sensing of Environment on Aug. 28.