Recent advances in the OCRA/ROCINN cloud retrieval framework for geostationary observations

The OCRA/ROCINN algorithm tandem is a mature and well-established framework for the retrieval of cloud macrophysical properties from passive satellite UV-VIS-NIR observations. OCRA (Optical Cloud Recognition Algorithm) derives the radiometric cloud fraction using image analysis, while ROCINN (Retrieval Of Cloud Information using Neural Networks) complements this information by retrieving two additional cloud macrophysical parameters through an optimal estimation approach supported by neural network emulation of radiative transfer simulations. Over the past decades, the OCRA/ROCINN algorithm tandem has formed the basis of the operational cloud products for a wide range of low Earth orbit missions, including GOME/ERS‑2, GOME‑2 on MetOp‑A/B/C and TROPOMI/S5P. Furthermore, OCRA/ROCINN has also been applied successfully to the EPIC/DSCOVR mission, located at the Lagrangian point L1.

In this contribution, the OCRA/ROCINN framework is considered in the context of geostationary cloud remote sensing, where its sub-daily temporal resolution leads to time-dependent scene illumination conditions. In recent years, the launch of new geostationary missions (GEMS, launched in February 2020; TEMPO, launched in April 2023; and Sentinel‑4, launched in July 2025) has established the Geo‑Ring constellation, providing coordinated observations for atmospheric composition and air quality monitoring over the Northern Hemisphere. In particular, OCRA/ROCINN constitutes the operational cloud processing algorithm tandem for Sentinel‑4, which motivates its assessment and further development in a geostationary setting alongside the other Geo‑Ring missions.

When applying the OCRA/ROCINN framework to geostationary missions, different practical challenges arise for its individual components. For OCRA, a key requirement is the availability of reference cloud-free maps, which typically relies on the accumulation of approximately one year of mission data. While this challenge is already present for low Earth orbit missions, where cloud-free maps can be supported by measurements from other low Earth orbit missions with similar local overpass times and wavelength coverage, it becomes significantly more restrictive in the geostationary case, where cloud-free maps are needed at sub-daily temporal resolution. Within the Geo‑Ring constellation, the lack of spatial overlap between the individual missions prevents mutual support, effectively limiting viable early-mission solutions to complementary observations from EPIC, which provides multiple daytime observations of all Geo‑Ring regions thanks to its location at the Lagrangian point L1. For ROCINN, the main challenge relates to the design of neural network architectures that emulate the radiative transfer model with high accuracy while remaining computationally efficient. Current developments focus on improving the representation of the forward model through suitable neural network configurations, for example by constraining the effective dimensionality of the neural network output space. This contribution focuses on these complementary advances within the OCRA/ROCINN framework in the context of geostationary cloud remote sensing.