New Algorithm Unlocks 10-Minute Ocean Color Observations from Himawari-8
08 Jul 2026
A new machine learning algorithm has enabled the Himawari-8 geostationary meteorological satellite to observe ocean color every 10 minutes, giving scientists a much sharper view of how coastal waters change throughout the day. By learning from high-quality MODIS ocean color observations, the transformer-based model overcomes key limitations of Himawari-8, including its relatively low signal-to-noise ratio, and significantly reduces retrieval errors compared with standard hourly products. The approach opens a new path for near-real-time monitoring of algal blooms, sediment transport, water quality, and other fast-changing ocean processes across the Asia-Pacific region.
Ocean color remote sensing is essential for observing marine ecosystems, primary productivity, coastal water quality, and harmful algal blooms. Polar-orbiting satellites, such as the Moderate Resolution Imaging Spectroradiometer (MODIS), provide accurate ocean color observations, but they usually pass over the same region only once or twice a day. As a result, rapid changes in dynamic coastal waters are often missed.
Geostationary satellites offer a complementary advantage. Positioned above the same region, Himawari-8 can observe the Earth at a very high temporal frequency. However, it was designed as a meteorological satellite, not a dedicated ocean color sensor. Its relatively low signal-to-noise ratio and the use of standard hourly composites can introduce systematic biases, including underestimation of remote sensing reflectance, or Rrs, in turbid waters and overestimation in clearer waters.
To address this challenge, researchers from the Aerospace Information Research Institute of the Chinese Academy of Sciences (AIRCAS), Inner Mongolia Normal University, the University of Oslo, and other partner institutions developed a transformer-based algorithm to retrieve Rrs from Himawari-8 multispectral data at 10-minute resolution. The study was published in the Journal of Remote Sensing on May 15, 2026.
The algorithm combines a conventional atmospheric correction framework with a transformer neural network. After correcting for gas absorption and Rayleigh scattering, the model learns the nonlinear relationship between top-of-atmosphere reflectance and ocean surface reflectance. Its inputs include solar viewing geometry, six Himawari-8 reflectance bands, aerosol optical thickness, and wind speed. High-quality MODIS Aqua Rrs products and AERONET-OC in situ measurements were used as reference data for training and validation.
Validation results show that the new algorithm outperforms official Himawari-8 Level-3 hourly products across all visible bands. Compared with AERONET-OC in situ observations, root mean square errors were reduced by 34%, 26%, and 12% at 470, 510, and 640 nm, respectively. The transformer model achieved correlation coefficients above 0.98 on test data, higher than both the random forest baseline and the operational product. Comparisons with MODIS ocean color products also showed strong spatial and temporal consistency, with correlation coefficients above 0.96.
The model also corrects known retrieval biases in different water types. It reduces the underestimation of Rrs at 470 and 510 nm in turbid coastal waters, while addressing overestimation at 640 nm in clearer waters. These improvements allow the system to capture rapid optical changes within a one-hour window, changes that are often smoothed out or missed by standard hourly composites.
“Himawari-8 was not built for ocean color, but our machine learning approach helps compensate for its hardware limitations,” said Prof. SHI Chong, corresponding author of the study. “By learning from the high-quality MODIS record, the transformer can effectively correct sensor noise and retrieval biases. This allows us to track coastal water optics at 10-minute intervals, which is valuable for monitoring fisheries, algal blooms, and water quality.”
The study demonstrates, for the first time, that accurate 10-minute Rrs retrieval is possible from a geostationary meteorological satellite. In effect, the algorithm turns Himawari-8 from a weather observer into a high-frequency ocean color monitor, providing a new tool for tracking coastal dynamics across the Asia-Pacific region.
Looking ahead, the team plans to expand the training dataset, improve the model's generalizability under different water and atmospheric conditions, and incorporate sun-glint correction. The framework may also be applied to other geostationary satellites, such as GK-2A and FY-4, and could support future high-frequency ocean color products by combining geostationary, polar-orbiting, and hyperspectral satellite observations.
Research News
New Algorithm Unlocks 10-Minute Ocean Color Observations from Himawari-8
A new machine learning algorithm has enabled the Himawari-8 geostationary meteorological satellite to observe ocean color every 10 minutes, giving scientists a much sharper view of how coastal waters change throughout the day. By learning from high-quality MODIS ocean color observations, the transformer-based model overcomes key limitations of Himawari-8, including its relatively low signal-to-noise ratio, and significantly reduces retrieval errors compared with standard hourly products. The approach opens a new path for near-real-time monitoring of algal blooms, sediment transport, water quality, and other fast-changing ocean processes across the Asia-Pacific region.
Ocean color remote sensing is essential for observing marine ecosystems, primary productivity, coastal water quality, and harmful algal blooms. Polar-orbiting satellites, such as the Moderate Resolution Imaging Spectroradiometer (MODIS), provide accurate ocean color observations, but they usually pass over the same region only once or twice a day. As a result, rapid changes in dynamic coastal waters are often missed.
Geostationary satellites offer a complementary advantage. Positioned above the same region, Himawari-8 can observe the Earth at a very high temporal frequency. However, it was designed as a meteorological satellite, not a dedicated ocean color sensor. Its relatively low signal-to-noise ratio and the use of standard hourly composites can introduce systematic biases, including underestimation of remote sensing reflectance, or Rrs, in turbid waters and overestimation in clearer waters.
To address this challenge, researchers from the Aerospace Information Research Institute of the Chinese Academy of Sciences (AIRCAS), Inner Mongolia Normal University, the University of Oslo, and other partner institutions developed a transformer-based algorithm to retrieve Rrs from Himawari-8 multispectral data at 10-minute resolution. The study was published in the Journal of Remote Sensing on May 15, 2026.
The algorithm combines a conventional atmospheric correction framework with a transformer neural network. After correcting for gas absorption and Rayleigh scattering, the model learns the nonlinear relationship between top-of-atmosphere reflectance and ocean surface reflectance. Its inputs include solar viewing geometry, six Himawari-8 reflectance bands, aerosol optical thickness, and wind speed. High-quality MODIS Aqua Rrs products and AERONET-OC in situ measurements were used as reference data for training and validation.
Validation results show that the new algorithm outperforms official Himawari-8 Level-3 hourly products across all visible bands. Compared with AERONET-OC in situ observations, root mean square errors were reduced by 34%, 26%, and 12% at 470, 510, and 640 nm, respectively. The transformer model achieved correlation coefficients above 0.98 on test data, higher than both the random forest baseline and the operational product. Comparisons with MODIS ocean color products also showed strong spatial and temporal consistency, with correlation coefficients above 0.96.
The model also corrects known retrieval biases in different water types. It reduces the underestimation of Rrs at 470 and 510 nm in turbid coastal waters, while addressing overestimation at 640 nm in clearer waters. These improvements allow the system to capture rapid optical changes within a one-hour window, changes that are often smoothed out or missed by standard hourly composites.
“Himawari-8 was not built for ocean color, but our machine learning approach helps compensate for its hardware limitations,” said Prof. SHI Chong, corresponding author of the study. “By learning from the high-quality MODIS record, the transformer can effectively correct sensor noise and retrieval biases. This allows us to track coastal water optics at 10-minute intervals, which is valuable for monitoring fisheries, algal blooms, and water quality.”
The study demonstrates, for the first time, that accurate 10-minute Rrs retrieval is possible from a geostationary meteorological satellite. In effect, the algorithm turns Himawari-8 from a weather observer into a high-frequency ocean color monitor, providing a new tool for tracking coastal dynamics across the Asia-Pacific region.
Looking ahead, the team plans to expand the training dataset, improve the model's generalizability under different water and atmospheric conditions, and incorporate sun-glint correction. The framework may also be applied to other geostationary satellites, such as GK-2A and FY-4, and could support future high-frequency ocean color products by combining geostationary, polar-orbiting, and hyperspectral satellite observations.