Robust spectral optimization algorithms for retrieving inherent optical properties of global oceans

Remote sensing has become an important means for monitoring global ocean color, and various inversion methods have been developed in recent decades. The inversion accuracy of spectral optimization methods, which constitute a critical type of ocean color remote sensing algorithm, could be significantly affected by the errors associated with semianalytical models and specific inherent optical properties (SIOPs) of water constituents. In this study, this issue is addressed from an error post-correction perspective by exploring the correlation patterns among inherent optical properties (IOPs). Specifically, cross-regression models for the magnitude parameters associated with IOPs are established based on <i>in situ</i> data. Subsequently, the residual information obtained from the models is employed to construct a feasible region for each magnitude parameter in the spectral optimization process. Finally, the semianalytical inversion problem is reformulated as a spectral optimization task constrained by global correlation regularization (G-CoReg) models. To mitigate the impact of heteroscedasticity among the magnitude parameters, an inversion algorithm based on locally weighted correlation regularization (L-CoReg) models is further proposed. Quantitative inversion experiments carried out on the NOMAD in situ dataset and the SeaWiFS matchup dataset demonstrate that the accuracy of the newly proposed G-CoReg and L-CoReg algorithms is generally superior to that of existing spectral optimization algorithms in terms of various magnitude parameters and performance metrics. Importantly, under the influence of atmospheric correction errors, the new algorithms also demonstrate greater robustness with lower accuracy degradation compared to algorithms without correlation regularization models.