Two-Stage Sea Level Anomaly Prediction Framework with Dual-Branch Feature Representation
Year: 2025
Type: article
Abstract: Sea Level Anomaly (SLA) is commonly employed to identify mesoscale ocean eddies, and its spatiotemporal prediction is useful for analyzing oceanic dynamic processes. Traditional SLA prediction models often operate directly on raw SLA data, which hinders the extraction of feature representations under complex oceanic dynamics, thereby limiting their predictive accuracy. To overcome this limitation, this paper proposes a two-stage SLA prediction framework with dual-branch feature representation, including a reconstruction stage and a prediction stage. In the first stage, a Dual-Branch Encoder (DB-Encoder) and a Spatially Adaptive Normalization Decoder (S-Decoder) are designed to extract latent feature representations. The DB-Encoder processes raw SLA data and one-hot encodings of iso-valued regions through two separate branches to enhance topological awareness. Then, Sparse regression based Vector Quantization (SVQ) is employed to represent features as weighted combinations of codebook vectors, effectively reducing noise and enhancing robustness. The S-Decoder further refines the reconstruction by applying spatially adaptive normalization, making the output more consistent with realistic ocean dynamics. During the second stage, features are extracted from each frame using the pretrained DB-Encoder and codebooks, and then passed to spatiotemporal prediction networks. The S-Decoder subsequently reconstructs these predicted features into the original spatial representation. Results from the South China Sea show that the proposed method improves both the reconstruction and prediction quality of SLA, while also achieving improved performance in eddy detection tasks, validating its effectiveness and potential for practical applications.
Authors: Kefan Wang, Qi Kang
Institution: Tongji University
Language: English
Subfield: Oceanography
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: diamond
Award: 面向连续任务的进化迁移学习方法研究
OpenAlex ID: W7117483622