A Two-Stage Spatiotemporal Prediction Framework for Sea Level Anomaly

Mesoscale ocean eddies are key oceanic processes and are typically identified using closed contours of Sea Level Anomaly (SLA). Therefore, accurate prediction of SLA is of great importance. However, traditional prediction models often directly process SLA data, making it difficult to effectively capture the topological characteristics of closed contours. In addition, these models often exhibit limited feature representation capabilities and robustness when handling complex physical processes. To address this problem, this paper proposes a two-stage SLA prediction framework that decomposes the SLA prediction task into representation learning and spatiotemporal prediction. In the first stage, a dual-branch encoder (DB-Encoder) is designed to process both the raw SLA data and the one-hot encoding generated from isovalue regions separately. Then, sparse regression-based vector quantization (SVQ) is performed through their respective codebooks, effectively representing local characteristics via weighted combinations. Furthermore, spatially adaptive normalization decoder (S-Decoder) is employed to constrain the decoding process, ensuring that the reconstructed results better conform to actual oceanic patterns. In the second stage, the trained DB-Encoder and SVQ extract feature frame by frame, which are then fed into a spatiotemporal prediction model to predict SLA over time. Finally, the S-Decoder maps the predicted results back to the original spatial domain. Experimental results in the South China Sea region show that the model effectively improves the SLA prediction accuracy and provides a novel and effective solution for ocean process forecasting.