Abstract
Satellite ocean color chlorophyll-a (Chl-a) records are heavily gapped due to clouds and orbital sampling, which can bias trend and phenology analyses and hinder studies of mesoscale dynamics. We present generalized data complementation model-empirical orthogonal function (GDCM-EOF), a self-supervised spatiotemporal gap-filling framework that learns to reconstruct missing Chl-a directly from partially observed satellite records, avoiding reliance on external reanalysis labels that are often unavailable or scale-mismatched for ocean color. The model combines (a) geolocation encoding to represent large-scale spatial gradients, (b) an EOF (empirical orthogonal function)-based low-rank modality enhancement that injects dominant-mode priors for background consistency, and (c) a stage-wise knowledge-transfer strategy with multiscale attention to improve robustness under extreme missingness. We train and evaluate on European Space Agency Ocean Colour Climate Change Initiative v6.0 daily Chl-a (1998 to 2017), using 1998 to 2014 for self-supervised training/validation and 2015 to 2017 as an independent test period. Experiments are conducted at 2 scales: a global 1° × 1° product over 60°S–60°N and a regional Caribbean Sea product at 0.04° × 0.04°. Using a withheld-observation protocol (15% withheld; 85% as input), GDCM-EOF consistently outperforms the classical data-interpolating empirical orthogonal function baseline: In the Caribbean test, root mean square error decreases from 0.107 to 0.019 and
R
2
increases from 0.803 to 0.998; globally, root mean square error decreases from 0.110 to 0.066 and
R
2
increases from 0.904 to 0.979. The reconstructions preserve large-scale background fields while recovering mesoscale structures such as tropical instability waves and estuarine–current interactions. Comparisons with the Copernicus-GlobColour gap-free product further support the physical plausibility of reconstructed patterns and enable improved analyses of spatial variability and long-term Chl-a trends.