Vision Transformer-Based Multi-Physics Field Prediction for Real Marine Environments

This paper implements the fine-tuning of a multiphysical field prediction model for the real marine environment based on a pre-trained model of atmospheric physical fields, and proposes optimisation ideas and fine-tuning strategies for more downstream tasks. The physical field of marine environment is affected by its own and atmospheric factors, and it is difficult to capture the deep features of the marine physical field for small area prediction and prediction with few input types. Simple networks such as RNN and LSTM can hardly meet the needs of complex data and increasing number of downstream tasks, and the Transformer has the powerful ability of deep feature extraction and task expansion, and this paper introduces the VisonTransformer architecture for atmospheric physical field prediction. In this paper, we introduce the data methodology and performance evaluation indexes for fine-tuning the atmospheric physical field pre-training model using VisonTransformer structure to predict the multi-physical field of the real marine environment in the Chinese region. The results show that this paper achieves higher prediction density of multi-physical field prediction of the marine environment in the Chinese region, with stable performance in short-medium- and long-term prediction, and significantly better performance than the traditional prediction methods in long-term prediction, which provides a pretraining substrate for the downstream tasks of spatial prediction in the Chinese region.