Improved K-SVD with Rate-Distortion Sparsity Control and EOF Seeding for Compressing High-Resolution Ocean Sound Speed Profiles of CORA2
Year: 2026
Type: article
Abstract: Sound speed profile (SSP) data in high-resolution ocean reanalysis products are voluminous, and compressed storage is a prerequisite for efficient distribution. The empirical orthogonal function (EOF) representation remains the standard low-dimensional description, and dictionary learning represents SSPs with fewer coefficients. However, most existing comparisons rest on coefficient counts rather than on the bytes actually required at the decoder. Using one-year CORA2 SSPs at two locations with 2920 profiles, this paper establishes a unified evaluation protocol in which every decoder-side quantity is converted into byte counts and evaluated after single-precision storage, with training, validation, and test data strictly separated. Within this protocol, an improved framework, the depth-regularized rate-distortion K-singular value decomposition (DRD-KSVD), is developed. Its rate-distortion-driven orthogonal matching pursuit (RD-OMP) coder accepts an atom only when the error reduction outweighs its byte cost, linking per-profile adaptive sparsity to the byte budget, and the rate penalty admits an effective upper bound. Training combines a depth-curvature regularization, closed-form atom updates, EOF-seeded initialization, decoupled training and encoding sparsities, and a uniform-weight objective. In the medium-to-high storage-rate range, DRD-KSVD dominates the original K-SVD in the rate-distortion sense, with errors lower by roughly one fifth to three tenths. EOF retains the outermost front, which is attributed to the nearly low-rank, single-mode-dominated, and shape-varying structure of the data. The improved framework and conclusions provide a reproducible quantitative basis for selecting compression methods for high-resolution SSP data.
DOI: 10.3390/jmse14222170
Authors: Wenbin Xiao(Conceptualization, Investigation, Software, Supervision, Visualization, Writing – review & editing), Siyuan Liao(Formal analysis, Software, Writing – original draft), Chengshen Xiao(Validation, Writing – original draft), Yongxian Wang(Funding acquisition, Methodology, Resources, Supervision, Validation, Writing – review & editing)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7234405863
