Título: Reconstructing High -Resolution Ocean Currents via Multistage Technique Applied to a Multiscale Neural Network
Autores: Reza Arefidamghani, Hamidreza Anbarlooei
Resumo: High-resolution (HR) ocean current data are crucial for resolving submesoscale structures, turbulent energy cascades, and other fine-scale processes that drive oceanic physical and biogeochemical dynamics. However, generating HR outputs using numerical models like the Hybrid Coordinate Ocean Model (HYCOM) is computationally intensive, particularly over large spatial domains or extended time periods. Additionally, satellite and in-situ observations often lack the spatial and temporal resolution necessary to directly capture these features.
To address these limitations, this study investigates deep learning-based super-resolution (SR) techniques as a cost-effective post-processing solution for reconstructing HR ocean surface velocity fields from low-resolution (LR) numerical simulations. Using synthetically downsampled HYCOM simulations over the Gulf of Mexico as a testbed, we adopt the Downsampled Skip-Connection Multi-Scale (DSC/MS) network as a baseline and benchmark its performance against traditional interpolation methods such as bilinear and bicubic upsampling.
To enhance reconstruction quality, two improvement strategies have been propose: Residual Learning, where the network learns to predict high-frequency residuals that are added to the initial SR output, and Multistage Refinement, where successive SR models iteratively improve the predictions to recover finer-scale structures.
Model performance is assessed using standard image-quality metrics (MSE, PSNR, SSIM, FSIM) and physics-informed di- agnostics, including spectral energy distribution and turbulent kinetic energy (TKE) recovery. Results show that the multistage refinement strategy significantly outperforms both interpolation and single-stage SR models, yielding reconstructions that are not only visually sharper but also more physically consistent.
This study highlights the potential of deep super-resolution architectures as efficient tools for enhancing ocean model out- puts, enabling more accurate and physically meaningful HR reconstructions without the computational cost of running full- resolution simulations.
Palavras-chave: super-resolution; ocean currents; deep learning; HYCOM; multistage refinement; residual learning; turbulent kinetic energy; spectral energy.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1175784
Artigo em PDF: CBIC_2025_paper1175784.pdf
Arquivo BibTeX:
CBIC_2025_1175784.bib