This computation involves the Level 2 processing algorithm applied to raw TROPOspheric Monitoring Instrument (TROPOMI) data.
Optical Cloud Recognition Algorithm (OCRA) is the S5P_CLOUD_OCRA heritage. In OCRA, optical sensor measurements are divided into two components: a cloud-free background and a remainder expressing the influence of clouds. OCRA was first developed for GOME in the late 1990s, when enough data from the three sub-pixel broad-band PMDs (Polarization Measurement Devices) had accumulated to allow for the construction of the global cloud-free composite which is the key element in the algorithm. Over the course of the 16-year GOME record, the
algorithm was refined and the cloud-free composite adjusted as more data became available. OCRA has also been applied to SCIAMACHY and GOME-2. Initial cloud-free composites for these sensors were based on GOME data before dedicated measurements became available from SCIAMACHY and GOME-2. For S5P_CLOUD_OCRA, the initial cloud-free composite will be based on GOME-2 and OMI (see section 5.2). Retrieval of Cloud Information using Neural Networks (ROCINN) is the S5P_CLOUD_ROCINN heritage. ROCINN is based on the comparison of measured and simulated satellite sun-normalized radiances in and near the O2 A-band, and it uses a neural network algorithm to retrieve cloud-top height and cloud-top albedo. ROCINN uses the cloud fraction input from OCRA as one starting point. Early versions of ROCINN used a transmittance model to compute simulated radiances, but the latest versions are based on the use of the VLIDORT radiative transfer scattering model.
For GOME and GOME-2, ROCINN Version 2.0 is the current operational algorithm in the GDP [GOME Data Processor]. This version is based on the assumption that clouds are simply Lambertian reflecting surfaces so the two main retrieval products are the cloud-top height and the cloud-top albedo itself. This is the “clouds-as-reflecting-boundaries” (CRB) model; see for example [van Roozendael et al., 2006] for GOME and [Loyola et al., 2011] for GOME 2.
Although ROCINN 2.0 is the heritage algorithm, there is an important point of departure for S5P. For TROPOMI/S5P, ROCINN Version 3.0 was initially used, which is based on a more realistic treatment of clouds as optically uniform layers of light-scattering particles (water droplets). This is the “clouds-as-layers" (CAL) model – here, the two main retrieval products are the cloud-top height and the cloud optical thickness. Details of this algorithm prototype may be found in [Schuessler et al., 2014]. Although the CAL model will be the default for S5P, it has been requested that the CRB method should also be retained as an option. CAL is the preferred method for the relatively small TROPOMI/S5P ground pixels (5.5 x 3.5
km2). The CRB approach works best with large pixels such as those from GOME (footprint 320 x 40 km2). [Schuessler et al., 2014] has shown that for the smaller GOME-2 pixels, CAL retrieval produces more reliable cloud information than that from CRB, not only with regard to the accuracy of the cloud parameters themselves but also with regard to the effect of cloud parameter uncertainties on total ozone accuracy. In OCRA, the intensity is regarded as a linear function of the radiometric cloud cover, and in
ROCINN, TOA radiances for partially cloudy scenarios are computed using a linearly weighted mean of the clear-sky and fully-cloudy calculations, the weighting factor being the cloud fraction. In the context of this IPA model, the two algorithms are consistent. With the notably smaller pixel size that comes with higher spatial resolution, 3-D cloud radiative effects will become an important consideration in error budgeting for the cloud algorithms. For more information on the processing chain please see the ATBD document.