New Satellite Algorithm Boosts Real-Time Extreme Rainfall Monitoring from FY-4B Satellite
22 Jul 2026
A new machine-learning algorithm could help meteorologists detect heavy rainfall more accurately and in near real time using China's Fengyun-4B geostationary meteorological satellite.
Developed by a research team led by Associate Professor JI Dabin from the Aerospace Information Research Institute of the Chinese Academy of Sciences (AIRCAS), in collaboration with Associate Professor BAO Shanhu from Inner Mongolia Normal University and XIE Yanhui from the Institute of Urban Meteorology, China Meteorological Administration, the algorithm addresses two persistent challenges in satellite-based precipitation monitoring: limited information beneath cloud tops and a severe imbalance between rainy and non-rainy samples.
Geostationary satellites continuously observe the same region of Earth, making them particularly valuable for tracking rapidly developing storms. However, most satellite rainfall algorithms rely heavily on temperatures measured at cloud tops. They cannot directly observe the atmospheric structure below the clouds, where water vapor, vertical air motion and other processes ultimately determine how much rain reaches the ground.
To overcome this limitation, the researchers incorporated CMA-BJ v2.0, a high-resolution atmospheric reanalysis dataset produced at a spatial resolution of three kilometers. Compared with the commonly used ERA5 reanalysis dataset, which has a resolution of about 31 kilometers, the finer-scale data provide a more detailed representation of water vapor and atmospheric motion beneath clouds.
The team also introduced a dynamic sample-balancing method. In rainfall datasets, dry observations greatly outnumber rainy ones, while extreme rainfall events are rarer still. As a result, machine-learning models often become biased toward weak rainfall and underestimate intense precipitation.
The new algorithm, named GSPE-DML, continuously adjusts the proportion of training samples representing different rainfall intensities. It also modifies its sampling strategy as seasons and weather systems change, allowing the model to adapt to varying meteorological conditions.
Tests showed that GSPE-DML maintained relatively low bias and stable correlations across different rainfall intensities. Its accuracy improved as rainfall became heavier, suggesting that the balancing strategy strengthened the model's ability to recognize extreme precipitation.
The researchers further tested the algorithm during the landfall of Super Typhoon Doksuri in 2023. GSPE-DML reproduced the intense rainfall band surrounding the typhoon eyewall as well as weaker rainfall in the outer spiral bands. The retrieved patterns closely matched observations from ground-based rain gauges and the microwave radiometer AMSR2.
Comparison of rainfall patterns during Typhoon Doksuri at 09:00 (UTC) on July 28, 2023. The GSPE-DML estimate (a) is compared with FY-4B QPE (b), APCP (c), GSMaP-NRT (d), AMSR2-L2 (e), and rain-gauge observations (f). (Image by AIRCAS)
Across multiple seasons, the algorithm achieved Heidke Skill Scores ranging from 0.52 to 0.62 and maintained relatively low false-alarm rates. It also outperformed several operational and near-real-time precipitation products in identifying rainfall areas, estimating rainfall intensity and preserving spatial and temporal consistency.
By combining higher-resolution atmospheric information with adaptive machine learning, the approach may improve rainfall monitoring in the East Asian monsoon region, where rapidly developing storms frequently bring flash floods and other weather-related hazards.
Research News
New Satellite Algorithm Boosts Real-Time Extreme Rainfall Monitoring from FY-4B Satellite
A new machine-learning algorithm could help meteorologists detect heavy rainfall more accurately and in near real time using China's Fengyun-4B geostationary meteorological satellite.
Developed by a research team led by Associate Professor JI Dabin from the Aerospace Information Research Institute of the Chinese Academy of Sciences (AIRCAS), in collaboration with Associate Professor BAO Shanhu from Inner Mongolia Normal University and XIE Yanhui from the Institute of Urban Meteorology, China Meteorological Administration, the algorithm addresses two persistent challenges in satellite-based precipitation monitoring: limited information beneath cloud tops and a severe imbalance between rainy and non-rainy samples.
The study was published in IEEE Transactions on Geoscience and Remote Sensing.
Geostationary satellites continuously observe the same region of Earth, making them particularly valuable for tracking rapidly developing storms. However, most satellite rainfall algorithms rely heavily on temperatures measured at cloud tops. They cannot directly observe the atmospheric structure below the clouds, where water vapor, vertical air motion and other processes ultimately determine how much rain reaches the ground.
To overcome this limitation, the researchers incorporated CMA-BJ v2.0, a high-resolution atmospheric reanalysis dataset produced at a spatial resolution of three kilometers. Compared with the commonly used ERA5 reanalysis dataset, which has a resolution of about 31 kilometers, the finer-scale data provide a more detailed representation of water vapor and atmospheric motion beneath clouds.
The team also introduced a dynamic sample-balancing method. In rainfall datasets, dry observations greatly outnumber rainy ones, while extreme rainfall events are rarer still. As a result, machine-learning models often become biased toward weak rainfall and underestimate intense precipitation.
The new algorithm, named GSPE-DML, continuously adjusts the proportion of training samples representing different rainfall intensities. It also modifies its sampling strategy as seasons and weather systems change, allowing the model to adapt to varying meteorological conditions.
Tests showed that GSPE-DML maintained relatively low bias and stable correlations across different rainfall intensities. Its accuracy improved as rainfall became heavier, suggesting that the balancing strategy strengthened the model's ability to recognize extreme precipitation.
The researchers further tested the algorithm during the landfall of Super Typhoon Doksuri in 2023. GSPE-DML reproduced the intense rainfall band surrounding the typhoon eyewall as well as weaker rainfall in the outer spiral bands. The retrieved patterns closely matched observations from ground-based rain gauges and the microwave radiometer AMSR2.
Comparison of rainfall patterns during Typhoon Doksuri at 09:00 (UTC) on July 28, 2023. The GSPE-DML estimate (a) is compared with FY-4B QPE (b), APCP (c), GSMaP-NRT (d), AMSR2-L2 (e), and rain-gauge observations (f). (Image by AIRCAS)
Across multiple seasons, the algorithm achieved Heidke Skill Scores ranging from 0.52 to 0.62 and maintained relatively low false-alarm rates. It also outperformed several operational and near-real-time precipitation products in identifying rainfall areas, estimating rainfall intensity and preserving spatial and temporal consistency.
By combining higher-resolution atmospheric information with adaptive machine learning, the approach may improve rainfall monitoring in the East Asian monsoon region, where rapidly developing storms frequently bring flash floods and other weather-related hazards.