Sea Level Anomalies Prediction Via Two-Stage Spatiotemporal Transformer

This study successfully applies the spatiotemporal transformer-based model to sea surface height anomaly prediction task. An improved ResNet autoencoder, named SW-GD Autoencoder, is proposed based on Sliced-Wasserstein distance and Gradient Difference loss functions to extract spatial feature information. This autoencoder demonstrates a reduced complexity and training difficulty compared to GAN-based architectures while maintaining high data generation quality. Additionally, the VidHRFormer model is employed for inference, introducing a large receptive field spatial feed-forward network module that enhances information transmission between adjacent patches in the feature map, thereby improving the observability of prediction results. To simplify training, we used a two-stage approach: first training the autoencoder separately, then combining it with VidHRFormer for final training. Compared to state-of-the-art ConvLSTM models, our method yields superior results.