||This seminar will be divided in two parts. Firstly, I will present some recent results about a review paper I am preparing. It deals with the different methods we find in the data assimilation literature to jointly estimate Q and R. These error covariance matrices are crucial because they control the relative weights of the model forecasts and the observations in filtering methods. I will remind the different methods and present some numerical comparisons on toy-models. Secondly, I plan to present various applications of data-driven methods in geophysics, not especially for data assimilation. I will show some applications of the analog method and deep learning in environmental problems, e.g. the nowcasting of solar irradiance using geostationary satellites and the classification of oceanic and atmospheric phenomena using SAR images.