New AI-Based Method Enables Hourly Monitoring of GPP from Geostationary Satellites

 Abstract
Advancements in geostationary satellites allow monitoring of terrestrial photosynthesis at sub-daily scales, offering unprecedented opportunities for understanding vegetation productivity. However, at short temporal scales, photosynthesis is highly influenced by illumination conditions, particularly diffuse radiation (Ddif). Existing empirical models often overlook Ddif’s impact, leading to uncertainties in hourly gross primary productivity (GPP) mapping. We determined that incorporating Ddif effects into GPP modeling improves clear-sky GPP mapping using Himawari-8. We employed a light gradient boosting machine (LGBM) considering Ddif and compared it with other empirical models: parametric, regression, and machine learning. The LGBM outperformed others, achieving an R2 of 0.8146 and root mean square error of 2.848 μmol CO₂/m2/s against ground observations from 2020 to 2021. To investigate input variable contributions in LGBM predictions, we performed SHapely Additive exPlanation (SHAP) analysis. Results confirmed that aerosol optical depth (AOD) had a greater impact during morning and evening when Ddif influence increased due to solar path length. Hourly GPP maps over East Asia from 2020 to 2021 using the LGBM demonstrated that diurnal patterns differ by landcover, with variations observed in latitudinal profiles. This underscores the need to examine GPP spatial distribution at high frequency. We confirmed that the spatial distribution of AOD SHAP values varied over time, highlighting temporal dynamics of aerosol effects. Our findings demonstrate the necessity of GPP mapping using geostationary satellites and suggest various impact studies can use our proposed framework. This approach provides a valuable tool for understanding vegetation’s rapid response to atmospheric aerosols, contributing to more accurate ecosystem flux modeling.

Approximately 30% of global carbon dioxide (CO₂) emissions are removed by terrestrial vegetation through photosynthesis. Researchers affiliated with UNIST have unveiled an innovative artificial intelligence (AI) analysis technique that predicts this CO₂ uptake with high temporal resolution. This advancement is expected to significantly aid climate change mitigation efforts and the development of carbon-neutral policies.

Led by Professor Jungho Im from the Department of Earth Environmental Urban Construction Engineering at UNIST, the research team announced the development of an AI model capable of estimating daily gross primary production (GPP) at hourly intervals by utilizing high-frequency data from geostationary weather satellites.

GPP is a key indicator representing the amount of carbon actively absorbed by plants during photosynthesis, serving as a crucial metric for quantifying ecosystem carbon sequestration.

The newly developed model leverages 10-minute interval observations from the Himawari-8 geostationary satellite to predict GPP with high temporal accuracy. First author Sejeong Bae explained, “Unlike polar-orbiting satellites, which typically observe a given location only 1 to 4 times per day, our model benefits from more frequent data collection, enabling precise estimation of diurnal changes in photosynthesis.”

Figure 1. AI-Based Daily Photosynthesis Monitoring and Analysis of the Impact of Fine Dust

The model incorporates various meteorological data, including Aerosol Optical Depth (AOD)—a satellite-derived indicator reflecting the concentration of particulate matter such as fine dust. AOD influences the amount and quality of sunlight reaching the surface by absorbing or scattering solar radiation, thereby affecting photosynthetic activity.

To interpret how the model makes predictions, the researchers employed SHapley Additive exPlanations (SHAP), an explainable AI technique. Results revealed that AOD is the most influential factor during morning and evening hours when the solar angle is low. This finding aligns with the understanding that lower solar elevation increases the proportion of scattered light, making vegetation’s photosynthetic response more sensitive to atmospheric aerosols.

Professor Im commented, “Our approach can estimate the spatial and temporal dynamics of carbon absorption over East Asia at a 2 km resolution across 24 hours, making it a valuable tool for ecosystem carbon flux analysis, vegetation monitoring, and light environment-based carbon modeling.”

Their findings have been published in Remote Sensing of Environment (IF: 11.1), one of the leading international journals in remote sensing and environmental sciences, on June 1, 2025. The research was supported by the Korea Environment Industry and Technology Institute (KEITI) under the Ministry of Environment, as well as the Korean Agency for Infrastructure Technology Advancement (KAIA) through the Ministry of Land, Infrastructure and Transport (MOLIT).

Journal Reference
Sejeong Bae, Bokyung Son, Taejun Sung, et al., “Advancing hourly gross primary productivity mapping over East Asia using Himawari-8 AHI and artificial intelligence: Unveiling the impact of aerosol-induced radiation dynamics,” Remote Sens. Environ., (2025).