A deep learning super-resolution model to speed up computations of coastal sea states

Abstract

In this paper, the potential of a super-resolution technique is presented in the context of coastal wave forecasting. The method uses a neural network to predict a high-resolution spatial estimation of spectral wave parameters from a lower resolution numerical computation. In this particular example, one year of training data is sufficient to achieve satisfying accuracy for practical applications. The error of this method in reproducing the results of a high-resolution spectral model is an order of magnitude lower than the usual accuracy of spectral models. Simultaneously, it reduces the computation time by a factor of up to 50. Moreover, utilizing complementary training data of extreme events allows for a further improvement in accuracy. The study also shows that super-resolution is more accurate, albeit slower, than surrogate models, thus providing a trade-off solution between accuracy and speed. Overall, incorporation of the present approach into wave forecasting systems has the potential to rapidly generate “zoomed-in” areas of interest or to speed up ensemble forecasts without supplementary calculations at higher resolution.