A Machine Learning Algorithm for Cloud Detection Based on the CO2M Multi-Angular Polarimetric Satellite Measurements

The Earth’s atmosphere contains suspended particles and molecules with a wide range of characteristics. Their interaction with radiation (in both solar and thermal spectral regions) affects the transfer of energy as well as its spatial distribution in the atmosphere, affecting the weather at any moment and climate in the long term. Multi-Angular Polarimetric (MAP) observations have a great potential for quantifying the properties (e.g., size, concentration, etc.) of aerosol particles at a high accuracy. For this reason, a MAP is included on the Copernicus Carbon Dioxide Monitoring satellite mission (CO2M; intended launch date: 2026) to provide a correction of the light path to meet the mission’s stringent requirements for CO2 column retrievals. However, for both trace gas and aerosol retrievals it is also essential to filter out any cloud-contaminated measurements, because clouds strongly interact with radiation and cover between 60-70% of the Earth’s surface at any given time. This study presents an algorithm designed for detecting clouds based on the MAP instrument on CO2M. The algorithm is an adaptation of an approach that was newly developed at SRON Netherlands Institute for Space Research for the MAP instrument onboard the Polarisation and Anisotropy of Reflectances for Atmospheric Science coupled with Observations from a Lidar (PARASOL) platform (i.e., POLarization and Directionality of Earth Reflectances; POLDER) and we are working towards making it applicable to other MAP instruments. This algorithm consists of an Artificial Neural Network model that is trained based on synthetic measurements with realistic geometry, aerosol, and cloud inputs. The synthetic measurements correspond to a wide range of atmospheric conditions and were produced for using the Remote Sensing of Trace Gases and Aerosol Products (RemoTAP) forward radiative transfer model developed at SRON Netherlands Institute for Space Research. This algorithm is designed to predict the cloud fraction based on the observed multi-angular polarization and radiance data, plus the instrument specifications and the corresponding viewing- and solar- geometry parameters. Here we focus on the efficacy of the approach for the CO2M mission. Furthermore, the sensitivity of the algorithm’s performance as a function of instrument characteristics (e.g, viewing angles, wavelengths, accuracy) will be discussed.