Newswise — Remote sensing of ocean color is essential for assessing marine ecosystems, primary productivity, and algal blooms. Polar-orbiting satellites like the Moderate Resolution Imaging Spectroradiometer (MODIS) provide accurate Rrs but miss diurnal variations. Geostationary satellites such as Himawari-8 provide high temporal resolution but are not dedicated ocean color sensors. Low signal-to-noise ratio (SNR) and standard hourly synthesis result in a systematic bias of underestimation in murky waters and overestimation in clear waters. As a result, rapid minute-by-minute changes in coastal areas remain poorly captured. Based on these challenges, there is an urgent need to leverage machine learning to fuse multi-source observations and develop dedicated approaches to obtain reliable high-frequency remote sensing reflectance.Rrs) From Himawari‑8.
On May 15, 2026, researchers from the Chinese Academy of Sciences, Inner Mongolia Normal University, the University of Oslo, and other partners. Jdiary of Remote Ssensing. They developed a transformer-based algorithm that retrieves the following information: Rrs Obtaining 10-minute resolution from Himawari-8’s multispectral data is a first for this geostationary weather satellite. This method addresses a critical real-world problem: Himawari-8’s low SNR, which reduces the accuracy of ocean colors. Learning from high-quality MODIS observations significantly improves the algorithm Rrs This allows rapid optical changes in dynamic coastal waters to be reliably monitored.
The new algorithm outperforms the official Himawari-8 Level-3 time product across all visible bands. Validation against AERONET‑OC (Aerosol Robotic Network‑Ocean Color) field data shows 34%, 26%, and 12% root mean square error reductions at 470 nm, 510 nm, and 640 nm, respectively. The Transformer model achieved a correlation coefficient greater than 0.98 on the test data. This is significantly higher than the Random Forest baseline (0.95) and the operational product (0.84). The key is to correct systemic bias, or underestimation. Rrs It is overestimated at 470/510 nm in murky coastal waters and 640 nm in clear waters. Comparison with the MODIS ocean color product shows strong spatial and temporal consistency (R > 0.96). This is the first demonstration of accurate 10-minute Rrs retrieval from a geostationary weather satellite.
The algorithm integrates a classical atmospheric correction (AC) framework and a transformer neural network. First, gas uptake (O₃ and NO₂) is corrected using the ECMWF Reanalysis v5 (ERA5) reanalysis and Ozone Monitoring Instrument (OMI) data, and Rayleigh scattering is removed via a lookup table from the coupled atmosphere-ocean radiative transfer model. The transformer selected for the self-attention function learns the nonlinear relationship between upper atmosphere reflectance and Rrs. Input features include solar zenith angle, Himawari-8’s six reflectance bands (470-2257 nm), aerosol optical thickness (AOT), and wind speed. The training targets are high quality MODIS Aqua Rrs products (spectral interpolation to Himawari-8 band) and AERONET-OC field measurements. Approximately 475 million samples were collected (9:1 training to test split). This model obtains Rrs of 470, 510, and 640 nm with a resolution of 5 km. Validation results demonstrate that this transformer reduces the hourly product error by 4-12.5% and captures rapid changes in coastal Rrs within a one-hour window. This is a feature not visible in standard composite materials. This algorithm maintains stable performance across seasons and different AOTs.
“While Himawari‑8 was not built with ocean colors in mind, our machine learning approach compensates for the limitations of the hardware,” said corresponding author Dr. Chong Shi. “By learning from high-quality MODIS recordings, the transformer effectively ‘corrects’ for sensor noise and recovery bias. This enables daytime tracking of coastal water optics at 10-minute intervals, which is important for fisheries, algae blooms, and water quality management. ”
The research team first performed gas absorption (O3, NO2) and Rayleigh scattering corrections using ERA5 pressure, an OMI NO2 column, and a radiative transfer lookup table. Himawari-8 Level 1 upper atmospheric reflectance (bands 1-6) matched MODIS Aqua Rrs products and AERONET-OC field data within ±10 minutes. The transformer model with multi-head self-attention was trained on 427.5 million samples. Inputs include display geometry, AOT, and wind speed. Model performance was evaluated using mean bias error, root mean square error, and correlation coefficient.
The algorithm transforms weather satellites into powerful ocean color monitors, enabling near real-time observations of coastal dynamics, harmful algal blooms, and sediment movement across the Asia-Pacific region every 10 minutes. Future work will extend the training samples to increase generalizability, incorporate solar brightness correction, and apply the framework to other geostationary satellites (e.g., GK-2A, FY-4). Longer time series validation assesses seasonal stability. Ultimately, this approach could be extended to hyperspectral sensors like PACE to provide consistent high-frequency ocean color products from both geostationary and polar-orbiting platforms.
###
References
Toi
10.34133/Remote Sensing.1047
Original source URL
https://doi.org/10.34133/remotesensing.1047
Funding information
National Natural Science Foundation of China 42275145, U24A20605Chong Shi. National Natural Science Foundation of China 42025504Husi Letu.
About remote sensing journal
of remote sensing journal, An online-only open access journal published in collaboration with AIR-CAS that promotes interdisciplinary research in the theory, science, and technology of remote sensing, as well as in the geosciences and information sciences.
