An Empirically Calibrated Optical Mask Approach for Estuarine Turbidity Front Detection with AlphaEarth Embeddings and Sentinel-2 Spectral–Spatial Features

Estuarine turbidity fronts are narrow transition zones where suspended particulate matter concentrations change sharply. Their detection from remote sensing imagery remains challenging because conventional methods rely on empirical thresholds, are sensitive to mixed pixels, and often lack transferability and physical interpretability. Here, we evaluate the potential of foundation-model representations by integrating AlphaEarth 64-dimensional embeddings with Sentinel-2 spectral and multi-scale spatial features. A 229-dimensional feature set is constructed and fed into a two-step framework combining random forest classification with an empirically calibrated optical mask based on low red-band reflectance. The fused feature set achieves an overall accuracy of 91.2%, an F1 score of 87.5%, and a Kappa coefficient of 0.807, outperforming both spectral–spatial features alone and AlphaEarth embeddings alone. To elucidate the contribution mechanism of AlphaEarth embeddings, we conduct two complementary SHAP analyses: one evaluating each dimension’s direct contribution to front classification, and the other assessing its capacity to predict Sentinel-2 band reflectance. Only nine dimensions overlap between the respective top 20 lists, revealing a clear functional division within the embedding space—some dimensions primarily encode spectral reflectance information, while others encode spatial context, edge patterns, or topological structures that are not directly accessible from local spectral features. This division represents the added value of AlphaEarth beyond conventional optical data. The empirically calibrated optical mask reduces candidate frontal area by 59.09% in turbid estuaries and restores linear front morphology. However, leave-one-estuary validation yields F1 scores ranging from 0.33 to 0.84, substantially below the within-estuary score of 0.93, demonstrating limited cross-region transferability and challenging the assumption of domain invariance in foundation-model embeddings. These findings highlight both the value of fusing foundation-model representations with local spectral–spatial features and the critical need for domain-adaptation strategies to improve generalization across contrasting estuarine hydrodynamic regimes.