image: Self-supervised training and evaluation workflow. During training, a random subset of valid observations is used as input together with an observation mask; the loss is computed only on valid (observed) locations to avoid supervising on missing pixels. During testing, a disjoint subset of valid observations is withheld as pseudo-ground truth, and metrics are computed only at these withheld locations. All preprocessing statistics (e.g., normalization and empirical orthogonal function [EOF] basis) are derived from the training period only to prevent data leakage.
Credit: Journal of Remote Sensing
Clouds and incomplete satellite sampling leave major gaps in ocean-color chlorophyll-a (Chl-a) observations, limiting studies of marine ecosystems and climate variability. A new self-supervised framework, GDCM-EOF, learns directly from incomplete satellite records without requiring external gap-free labels. Compared with the DINEOF baseline, reconstruction root mean square error (RMSE) is about 82% lower in a high-resolution Caribbean Sea test and 40% lower in a global test, with both results assessed in log-transformed Chl-a space.
Satellite ocean-color records provide essential information on phytoplankton biomass, marine primary productivity, and carbon cycling. Yet clouds, unfavorable light conditions, and sensor coverage leave substantial gaps; the study reports a mean missing fraction of 85.3% in the examined OC-CCI dataset. These gaps can distort trend and phenology estimates and obscure fronts, eddies, and river plumes. Simple spatial interpolation can oversmooth local variability, while data-interpolating empirical orthogonal functions (DINEOF) favor dominant patterns and may miss fine-scale structures. Although supervised deep learning can reconstruct complex fields, it often requires external complete reference data. Suitable chlorophyll-a reanalyses are limited and may differ from satellite products in resolution, motivating reconstruction methods trained directly on partially observed records.
A research team from the Institute of Oceanology, Chinese Academy of Sciences, with the Center for Ocean Mega-Science and the University of Chinese Academy of Sciences, published a study (DOI: 10.34133/remotesensing.1064) in Journal of Remote Sensing on 7 September 2026. The team developed GDCM-EOF (generalized data complementation model–empirical orthogonal function) to address missing ocean-color observations. Rather than learning from externally interpolated reference fields, it uses observed pixels for self-supervised training. The study demonstrates reconstruction on a global 1° grid and a regional 0.04° Caribbean Sea grid, not a global 0.04° product.
In withheld-observation tests, GDCM-EOF outperformed the conventional DINEOF baseline. In the Caribbean Sea, RMSE fell from 0.107 to 0.019 (an 82% reduction) and R² rose from 0.803 to 0.998; globally, RMSE decreased from 0.110 to 0.066 (40%) and R² increased from 0.904 to 0.979. These metrics were computed in log10-transformed chlorophyll-a space, only at observed pixels deliberately withheld for evaluation. The reconstructions show coherent broad-scale patterns and finer structures associated with tropical instability waves, river plumes, and eddy activity. Comparison with Copernicus GlobColour supports the plausibility of these spatial patterns, but that gap-free product does not provide independent ground truth for every originally missing pixel.
The researchers used ESA Ocean Colour Climate Change Initiative (OC-CCI) v6.0 daily Chl-a records from 1998 to 2017. Data from 1998–2014 supported training and validation; 2015–2017 was held out for testing. Experiments were conducted at 1° × 1° over 60°S–60°N globally and 0.04° × 0.04° in the Caribbean Sea. For quantitative evaluation, 15% of pixels with valid observations were withheld as reference values, while the remaining 85% of valid pixels and the original missing-data pattern formed the input. Accuracy therefore describes agreement with withheld observations, not verified error across all cloud-obscured regions. The model integrates Fourier geolocation encoding, an EOF-based low-rank prior using the leading five modes, multiscale attention, and three stages of knowledge transfer. Preprocessing statistics and EOF bases were derived from the training period to prevent test-data leakage.
The central contribution is a way to learn from satellite observations that are incomplete from the outset. During training, only observed pixels provide supervision; during testing, another subset of genuinely observed pixels is held back to assess generalization. This distinction matters because model estimates at persistently unobserved locations cannot be directly verified by the same protocol. The authors also identify computational expense, extended observational gaps, and input-data quality as remaining challenges.
At the model level, geolocation encoding supplies latitude–longitude context, while an EOF block extracts the leading five spatial modes to provide preliminary low-rank estimates in missing areas. These estimates feed an encoder–decoder network with ConvLSTM, spatial–temporal attention, residual connections, and atrous spatial pyramid pooling. A stage-wise training strategy passes model weights and updated estimates from one stage to the next, refining reconstruction without using unknown pixels as target labels. Case studies reveal patterns associated with tropical instability waves, the Amazon River plume–North Brazil Current system, and filaments and eddies in the Caribbean Sea.
Future work will explore lighter architectures, model compression, and more efficient training to reduce the additional computational cost relative to DINEOF. Diffusion-based reconstruction and quantified uncertainty could help address ambiguity under extended cloud cover or other persistent gaps. The team also aims to investigate joint reconstruction of chlorophyll-a with sea surface temperature, salinity, and other biogeochemical variables to exploit cross-variable covariability. Such developments may support more continuous ocean-color records for ecosystem monitoring, carbon-cycle studies, and future operational applications, although independent validation and robustness to sensor inconsistencies and preprocessing artifacts will remain important.
###
References
DOI
Original Source URL
https://spj.science.org/doi/10.34133/remotesensing.1064
Funding information
This work is supported by the Self-Deployed Project of the Institute of Oceanology, Chinese Academy of Sciences (IOCASZZZX103), the National Natural Science Foundation of China (42606251, 42306194, 42076200, and 42376175), Shandong Province Postdoctoral Innovation Programme SDCX-ZG-202601023 and the China Postdoctoral Science Foundation (grant nos. GZC20250594 and 2025M780841).
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.
Journal
Journal of Remote Sensing
Subject of Research
Not applicable
Article Title
Self-Supervised Spatiotemporal Reconstruction of Ocean Color Chlorophyll-a from Satellite Observations
Article Publication Date
7-Sep-2026
COI Statement
The authors declare that they have no competing interests.