Clouds cover approximately 60% of the globe and are therefore an obstacle for observing the Earth’s surface from space in a wide range of wavelengths. To limit their impact on Infrared Atmospheric Sounding Interferometer (IASI)-based atmospheric and surface property retrievals, it is essential to have access to accurate cloud properties. Most cloud retrievals, whether physical or statistical, are mostly performed at the pixel-level. Using the spatial coherency within the IASI footprint is however beneficial for the detection of clouds, that are spatially structured. We propose to use IASI orbits, restructured as rectangular images, collocated with the cloud classification (clear, water, ice, two-level ice) extracted from SEVIRI-based Optimal Cloud Analysis (OCA) to develop a global cloud classification product using Partial Convolutional Neural Networks (CNN). For the first time, the partial convolutional layer is introduced as a way to deal with the large amount of spatially missing data. The partial-CNN model correctly classifies the cloud type with 75% accuracy (this number increases to 87% when considering only spatially homogeneous IASI pixels). Important differences are mostly related to IASI/SEVIRI resolution discrepancy. Thanks to the image processing approach and the cloud spatial coherency, two-layer clouds are retrieved, which is not possible with a pixel-wise processing. Our product not only classifies the cloud phase globally but also estimates, using the network a posteriori class probabilities, the cloud type fractions covering each IASI pixel.

Le séminaire se tiendra au LMD, à l’École Polytechnique, dans la salle PMC, 2e étage.