Key points are not available for this paper at this time.
Abstract. Numerical forecast products with high temporal resolution are crucial tools in atmospheric studies, allowing for accurate identification of rapid transitions and subtle changes that may be missed by lower-resolution data. However, the acquisition of high-resolution data is limited due to excessive computational demands and substantial storage needs in numerical models. Current deep learning methods for statistical downscaling still require massive ground truth with high temporal resolution for model training. In this paper, we present a self-supervised framework for downscaling atmospheric variables at arbitrary time resolutions by imposing a temporal coherence constraint. Firstly, we construct an encoder–decoder-structured temporal downscaling network and then pretrain this downscaling network on a subset of data that exhibit rapid transitions and are filtered out based on a composite index. Subsequently, this pretrained network is utilized to downscale the fields from adjacent time periods and generate the field at the middle time point. By leveraging the temporal coherence inherent in meteorological variables, the network is further trained based on the difference between the generated field and the actual middle field. To track the evolving trends in meteorological system movements, a flow estimation module is designed to assist with generating interpolated fields. Results show that our method can accurately recover evolution details superior to other methods, reaching 53.7 % in the restoration rate on the test set. In addition, to avoid generating abnormal values and to guide the model out of local optima, two regularization terms are integrated into the loss function to enforce spatial and temporal continuity, which further improves the performance by 7.6 %.
Wang et al. (Thu,) studied this question.