Sea Level Anomaly prediction with TSTA-enhanced UNet

Abstract

The prediction of Sea Level Anomaly (SLA) is crucial for many applications in marine and meteorological tasks. Most recently developed SLA prediction methods have been developed mainly on the framework of the Recurrent Neural Network (RNN) and its variants. These frameworks suffer from insufficient capability to capture spatial information and low computational efficiency. To address these issues, this paper proposes a novel method called UNet and Temporal-Spatial Transformer Attention (UNet-TSTA) for accurate and efficient SLA prediction. In our model, UNet serves as the backbone structure of the prediction model, enhancing the model’s ability to capture features of sea surface eddies at different scales. Meanwhile, the TSTA module innovatively constructs multiple spatial–temporal planes through the free combination of temporal and spatial dimensions, utilizing the attention mechanism of the Point-by-Point Vision Transformer (P-ViT). The effective cooperation of P-ViT and CNN also enhances the training and inference speed of the model. Experimental results on real SLA datasets show that our UNet-TSTA method achieves millimeter-level average precision in predicting SLA fields for the next seven days. Compared to other advanced algorithms, our method shows significant improvements in both computational efficiency and prediction precision.