Cross-Estuary Generalization of Color Front Identification Using DenseNet-121

Remote identification of color fronts, defined as transition zones with sharp gradients in water optical properties, has suffered from non-transferable thresholds, severe areal over-detection, and weak cross-estuary generalization. To address these issues, we propose a framework that integrates multi-scale spectral–spatial features with DenseNet-121. We constructed a 165-D vector, seven window scales (three × three to 15 × 15) × two statistical descriptors (means and standard deviations) × 11 bands + 11 bands, then rearranged it into a 3D tensor and resized it to a 2D image for DenseNet-121 transfer learning with red-band post-processing. On in-distribution tests, the model achieves 0.953 accuracy, 0.953 F1, outperforming random forest. Cross-estuary generalization yields a mean F1 (0.744). Performance varies with optical compatibility: the Mississippi (runoff-dominated) gives the best F1 (0.874), while the Pearl (multi-channel, runoff-tide co-controlled) drops to 0.607 due to heterogeneity and reversed reflectance patterns. The red-band constraint can help reduce areal false alarms and improve spatial coherence of frontal regions, but its effectiveness depends on optical separability. The output width reflects superposition of transition zone and window scale. We demonstrate the potential and boundary conditions of this approach for cross-estuary color front identification, offering insights for physically consistent and generalizable ocean color monitoring.