Satellite wave 2D spectrum partition based on the PI-vit-GAN(physically-informed ViT-GAN) method

Abstract

The abundant spectral data provided by satellite technology are crucial for interpreting the complex marine environment, and the effective and accurate analysis of these data is particularly important for coastal engineering. In this regard, this study proposes a Physically Informed ViT-GAN (PI-ViT-GAN) automatic partitioning method, based on CFOSAT satellite wave spectrum data. Specifically, the model consists of a generator and discriminator. The generator utilizes a contrastive learning strategy as pretraining and through the self-attention mechanism of the ViT model, it focuses on key parts of the spectrum to extract wave group features and wave element parameters. Partitioning-head joint training realizes the output of wave group partition element indices. Subsequently, the discriminator uses the wave group features and a parametric model for spectrum reconstruction and computes the error with the original observed spectrum to evaluate the partition and reconstruction effects. Additionally, this model incorporates two physically corrected functions, wave system classification loss and merging loss, based on the wave age criterion, thereby guiding the training process, and enhancing model efficiency. The results indicate that the reconstructed theoretical spectrum, obtained through the utilization of this method, aligns well with the original sea wave spectrum, demonstrating a precision superior to the spectral partitioning product of CFOSAT's own SWIM. Combining the robust learning capability of the transformer and the regularization of physical prior knowledge, this model can achieve precise, low-cost automated analysis of satellite wave spectra, providing a new scalable method for big data analysis in marine and coastal engineering.