Super-resolution data assimilation

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