A Multi-Element Joint Downscaling Method for Oceanography Based on Spatiotemporal Analysis and Deep Learning

Abstract

This invention, based on spatiotemporal analysis and deep learning, proposes a joint downscaling method for multiple oceanic elements, belonging to the field of meteorological and climate prediction and data processing technology. It includes: S1, data preparation and preprocessing; S2, multivariate empirical mode decomposition (MEOF); ​​S3, constructing an SRGAN model capable of learning spatial mode mapping relationships at different resolutions; and S4, generation and reconstruction of high-resolution ocean fields. This invention obtains joint spatial modes and time series through multivariate empirical mode decomposition. The model's learning of the joint spatial modes maintains the inherent spatiotemporal coupling relationships between multiple variables. It significantly improves the detail restoration capability of the downscaling results. In adversarial training, the model can generate more realistic and detailed high-resolution spatial structures, effectively capturing complex nonlinear processes in the climate system. It is highly efficient and flexible; after model training, it can quickly downscale long-term, multi-scenario low-resolution data, greatly saving computational resources.

Images