Development of an Inversion Method based on Convolutional Neural Networks for the retrieval of integrated water vapor above clouds in the context of the C3IEL mission.

Despite recent advances, modeling convective clouds remains an important source of uncertainties in climate and weather modeling. Their development is largely dependent on the amount of water vapor available in the atmosphere, which will increase with global warming. It is therefore necessary to better understand the spatial and temporal variability of water vapor in the atmosphere to improve our understanding of the interaction between this gas and clouds. To address this need, the C3IEL (Cluster for Cloud Evolution, Climate and Lightning) mission, a joint effort between CNES and ISA is developed and scheduled for 2028. This mission will use the differential absorption of water vapor in three Short-Wave infrared (SWIR) channels to retrieve the amount of integrated water vapor above and around convective clouds at a high spatial resolution of about 100 m. Recent studies have demonstrated the feasibility of using the optimal estimation method to perform such retrieval, based on the assumption of a plane-parallel cloud. However, despite accurate retrievals with RMSE less than 1kg /m², this method is computationally expensive and does not take into account the spatial context of the scene (pixel wise retrievals). This work presents another method based on convolutional neural networks – a computer vision deep learning architecture – to retrieve integrated water vapor above clouds and in clear sky areas. An attention mechanism and physical constraints are implemented to ensure the physical accuracy of the retrievals. The training of the presented model is based on synthetic C3IEL observations generated using the Meso-NH numerical atmospheric model and the ARTDECO 1D radiative transfer model. The first results are encouraging, with very fast retrievals inferior than 0.9 kg/m² RMSE on synthetic data and a real improvement brought by the attention mechanism and physical constraints. However, available training data are still limited due to computational costs of generating new cloudy scenes and new radiative transfer simulations, and current work aims to provide more diversity in training examples to really demonstrate the ability of the algorithm to generalize to new cases.