Radiometric calibration represents an indispensable foundation for quantitative ocean color remote sensing, with cross-calibration serving as a cornerstone methodology. However, existing cross-calibration methods primarily rely on linear calibration models. These approaches struggle to accurately characterize non-linear response residuals across complex environments and overlook both the spatial correlations among adjacent pixels and the spectral correlations between distinct bands. To address these challenges, this study departs from the conventional coefficient-fitting paradigm and proposes an innovative feature-driven radiometric cross-calibration method. By learning high-dimensional image features, this approach establishes a direct sensor-to-sensor mapping, obviating the need for explicit calibration coefficients. Specifically designed for the three-dimensional spatial-spectral characteristics of satellite imagery, the proposed Dual-Branch Spatial-Spectral Collaborative Network (DB-SSCN) synergizes local spatial features and global spectral dependencies, incorporating a band-weighted loss function to balance contributions across different bands. In a case study utilizing the Chinese Ocean Color and Temperature Scanner (COCTS) onboard the Haiyang-1D (HY-1D) satellite, the calibration results exhibit strong consistency with the standard radiance products of Aqua MODIS, yielding a Mean Squared Error (MSE) of 0.2661 and a Mean Relative Error (MRE) of 1.18%. Furthermore, the developed algorithm demonstrates exceptional spatial and spectral preservation capabilities, achieving a Structural Similarity Index Measure (SSIM) of 0.9879 and a Spectral Angle (SA) of 0.28°. The method effectively suppresses striping noise, thereby significantly improving the quality of the radiometric calibration products. Application experiments confirm that the method exhibits robust applicability across diverse water environments.