New AI Maps Plant Carbon Uptake Daily at 30-Meter Resolution
Abstract Satellite-based solar-induced chlorophyll fluorescence (SIF) has opened a new era in monitoring gross primary productivity (GPP), and near-infrared reflectance of vegetation (NIRv), an effective proxy for SIF, has enabled observations at high spatial resolution. However, the inherent trade-off between spatial and temporal resolution in remote sensing often leads to the loss of critical information on carbon cycling, particularly in heterogeneous and cloud-prone regions. To address this challenge, this study introduces the unified, high-resolution intelligent carbon quantification and explanation (UNIQUE) framework for high-spatiotemporal-resolution satellite-based GPP monitoring. UNIQUE comprises two components. Part 1 generates high-accuracy GPPL and GPPM from Landsat and the Moderate Resolution Imaging Spectroradiometer, respectively. Part 2 fuses these products to create GPPUNIQUE, which provides daily GPP estimates on a 30 m grid, with fine-scale spatial variability primarily derived from Landsat-based vegetation information. In Part 1, a light gradient boosting machine model using NIRvP, the product of NIRv and photosynthetically active radiation, as the key predictor achieved a root mean square error (RMSE) of 2.17–2.29 gC m-2 d-1, representing a 13–28% reduction relative to existing global GPP products. In Part 2, the proposed stochastic–deterministic combined architecture for fine-level regression (U-SCALER) achieved performance comparable to that of a deterministic UNET baseline. Its performance improved when available Landsat observations were incorporated, achieving an RMSE of 2.60 gC m-2 d-1 at test sites. U-SCALER also showed lower sensitivity to Part 1 errors and image-pair selection. Although meteorological drivers were incorporated at coarser spatial resolutions, GPPUNIQUE represents fine-scale spatial variability on a 30 m grid by leveraging Landsat-derived vegetation information. Overall, UNIQUE provides a practical approach for integrating satellite observations with different spatial and temporal characteristics. U-SCALER in Part 2 also contributes to the evaluation of generative modeling for image fusion and super-resolution. With potential extensions to multiple satellite sensors and to vegetation and carbon-related variables beyond GPP, UNIQUE could support more comprehensive carbon monitoring across diverse domains and spatiotemporal scales. Satellite observations can reveal how much carbon plants absorb through photosynthesis, but capturing these changes in both fine detail and at frequent intervals remains a challenge. Researchers at UNIST have developed an AI framework that produces daily carbon uptake estimates on a 30-meter grid, revealing differences that conventional satellite products may miss. Led by Professor Jungho Im of the Department of Civil, Urban, Earth, and Environmental Engineering, the team developed UNIQUE (Unified, High-Resolution Intelligent Carbon Quantification and Explanation) to overcome the trade-off between spatial detail and observation frequency in satellite-based carbon monitoring. Scientists estimate carbon uptake through photosynthesis using a measure called gross primary productivity (GPP). Satellites can monitor GPP across large areas, but those capturing finer details typically observe less frequently, while satellites providing daily data offer coarser views. Cloud cover adds another challenge. UNIQUE combines data from Landsat, which observes vegetation at a 30-meter resolution, and the Moderate Resolution Imaging Spectroradiometer (MODIS), which provides daily observations. Machine-learning models first estimate GPP separately from each source, using vegetation reflectance and environmental conditions that influence photosynthesis. The team's second AI model, U-SCALER, then merges these estimates into daily GPP maps on a 30-meter grid. Landsat supplies the fine-scale vegetation information, while the combined observations help capture changes over time. At Taehwa Mountain in South Korea, UNIQUE distinguished differences in carbon uptake between deciduous broadleaf and evergreen coniferous forests at sites just 320 meters apart. A conventional 500-meter-resolution product showed nearly identical patterns for both forests. The framework also captured variations across mixed farms and forests in Wisconsin, United States. In tests, the satellite-specific models reduced GPP estimation errors by **13–28%** compared with existing global products. The final estimates, evaluated against measurements from 30 ground-based monitoring sites, achieved a root mean square error (RMSE) of **2.60 grams of carbon per square meter per day**, approximately 8% lower than the established ESTARFM image-fusion method. “By extending this approach to additional satellite sensors and carbon-related variables, we hope to support more detailed monitoring of forests and farmland, assess how climate-related disasters affect ecosystems, and inform carbon-neutrality policies,” said Professor Im. Dr. Bokyung Son and Dr. Taejun Seong of UNIST served as co-first authors, with contributions from researchers at Seoul National University, Pusan National University, and the University of Freiburg in Germany. An optimized version of the method is already used in the Korea Carbon Project, a regional carbon information platform supported by the Korea Environmental Industry & Technology Institute (KEITI). The project is led by Professor Sujong Jeong of Seoul National University, with Professor Gayoung Yoo of Kyung Hee University overseeing terrestrial carbon uptake assessment. The research was supported by KEITI under the Ministry of Climate, Energy and Environment (MCEE), the National Research Foundation of Korea (NRF) under the Ministry of Science and ICT (MSIT), and the InnoCORE initiative. The findings were published online in Remote Sensing of Environment on August 28, 2026. Journal Reference Bokyung Son, Taejun Sung, Sejeong Bae, et al ., “UNIQUE: UNified, high-resolution Intelligent carbon QUantification and Explanation,” Remote Sens. Environ. , (2026).
2026.11.04