Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OptFormer: Optical Flow-Guided Attention and Phase Space Reconstruction for SST Forecasting

Yin Wang, Chunlin Gong, Zhuozhen Xu, Lehan Zhang, Xiang Wu

TL;DR

OptFormer tackles SST forecasting by unifying phase-space reconstruction with optical-flow-guided attention in an encoder–decoder framework, enabling data-efficient learning of nonlinear spatiotemporal dynamics. By embedding delayed attractors via Takens-like phase-space reconstruction and guiding attention with dense optical flow, the model concentrates on dynamically evolving regions and captures long-range dependencies. Empirical results on NOAA OISST data show significant improvements in RMSE and MAPE across spatial scales, horizons, and seasons, with ablations confirming the contributions of optical attention, multi-scale feature encoding, and temporal auto-correlation modeling. The work offers a practical, GPU-friendly approach for ocean monitoring and climate forecasting, while outlining future directions toward multivariate inputs and PDE-aware modeling.

Abstract

Sea Surface Temperature (SST) prediction plays a vital role in climate modeling and disaster forecasting. However, it remains challenging due to its nonlinear spatiotemporal dynamics and extended prediction horizons. To address this, we propose OptFormer, a novel encoder-decoder model that integrates phase-space reconstruction with a motion-aware attention mechanism guided by optical flow. Unlike conventional attention, our approach leverages inter-frame motion cues to highlight relative changes in the spatial field, allowing the model to focus on dynamic regions and capture long-range temporal dependencies more effectively. Experiments on NOAA SST datasets across multiple spatial scales demonstrate that OptFormer achieves superior performance under a 1:1 training-to-prediction setting, significantly outperforming existing baselines in accuracy and robustness.

OptFormer: Optical Flow-Guided Attention and Phase Space Reconstruction for SST Forecasting

TL;DR

OptFormer tackles SST forecasting by unifying phase-space reconstruction with optical-flow-guided attention in an encoder–decoder framework, enabling data-efficient learning of nonlinear spatiotemporal dynamics. By embedding delayed attractors via Takens-like phase-space reconstruction and guiding attention with dense optical flow, the model concentrates on dynamically evolving regions and captures long-range dependencies. Empirical results on NOAA OISST data show significant improvements in RMSE and MAPE across spatial scales, horizons, and seasons, with ablations confirming the contributions of optical attention, multi-scale feature encoding, and temporal auto-correlation modeling. The work offers a practical, GPU-friendly approach for ocean monitoring and climate forecasting, while outlining future directions toward multivariate inputs and PDE-aware modeling.

Abstract

Sea Surface Temperature (SST) prediction plays a vital role in climate modeling and disaster forecasting. However, it remains challenging due to its nonlinear spatiotemporal dynamics and extended prediction horizons. To address this, we propose OptFormer, a novel encoder-decoder model that integrates phase-space reconstruction with a motion-aware attention mechanism guided by optical flow. Unlike conventional attention, our approach leverages inter-frame motion cues to highlight relative changes in the spatial field, allowing the model to focus on dynamic regions and capture long-range temporal dependencies more effectively. Experiments on NOAA SST datasets across multiple spatial scales demonstrate that OptFormer achieves superior performance under a 1:1 training-to-prediction setting, significantly outperforming existing baselines in accuracy and robustness.
Paper Structure (24 sections, 24 equations, 4 figures, 5 tables)

This paper contains 24 sections, 24 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Physics-based models
  4. Data-driven models
  5. Methodology
  6. Problem Formulation
  7. Phase Space Reconstruction.
  8. Network Architecture
  9. Optical Attention
  10. Optical Flow Estimation
  11. Information Aggregation
  12. Experiment Design
  13. Dataset
  14. Data Construction
  15. Baselines
  16. ...and 9 more sections

Figures (4)

  • Figure 1: Overall architecture of OptFormer.
  • Figure 2: Structure of the Optical Attention module.
  • Figure 3: Temporal evolution of SST data with optical flow overlay. The first row displays the original SST field, optical flow magnitude, and their overlay at the initial time step (t=0). The second and third rows show the evolution of the overlay at subsequent time steps (t=1-6).
  • Figure 4: All models are evaluated at their respective converged epochs to ensure a fair comparison. As observed, OptFormer converges significantly faster than the baseline methods, indicating more efficient optimization and faster fitting to the underlying spatiotemporal dynamics.