Super-resolution data assimilation

• DOI: 10.1007/s10236-022-01523-x
• Publication Date: 2022-08-11T07:02:51Z
• Authors: Sébastien Barthélémy; Julien Brajard; Laurent Bertino; François Counillon
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; 0207 environmental engineering; FOS: Physical sciences; 02 engineering and technology; Computational Physics (physics.comp-ph); Nonlinear Sciences - Chaotic Dynamics; 530; 01 natural sciences; Machine Learning (cs.LG); Physics - Atmospheric and Oceanic Physics; Physics - Data Analysis, Statistics and Probability; Atmospheric and Oceanic Physics (physics.ao-ph); Chaotic Dynamics (nlin.CD); Physics - Computational Physics; Data Analysis, Statistics and Probability (physics.data-an); 0105 earth and related environmental sciences

AbstractIncreasing model resolution can improve the performance of a data assimilation system because it reduces model error, the system can more optimally use high-resolution observations, and with an ensemble data assimilation method the forecast error covariances are improved. However, increasing the resolution scales with a cubical increase of the computational costs. A method that can more effectively improve performance is introduced here. The novel approach called “Super-resolution data assimilation” (SRDA) is inspired from super-resolution image processing techniques and brought to the data assimilation context. Starting from a low-resolution forecast, a neural network (NN) emulates the fields to high-resolution, assimilates high-resolution observations, and scales it back up to the original resolution for running the next model step. The SRDA is tested with a quasi-geostrophic model in an idealized twin experiment for configurations where the model resolution is twice and four times lower than the reference solution from which pseudo-observations are extracted. The assimilation is performed with an Ensemble Kalman Filter. We show that SRDA outperforms both the low-resolution data assimilation approach and a version of SRDA with cubic spline interpolation instead of NN. The NN’s ability to anticipate the systematic differences between low- and high-resolution model dynamics explains the enhanced performance, in particular by correcting the difference of propagation speed of eddies. With a 25-member ensemble at low resolution, the SRDA computational overhead is 55% and the errors reduce by 40%, making the performance very close to that of the high-resolution system (52% of error reduction) that increases the cost by 800%. The reliability of the ensemble system is not degraded by SRDA.