Submesoscale horizontal density gradients at the ocean surface play a vital role in driving upper-ocean mixing and air-sea exchange. However, horizontal density structures are often assumed to directly reflect the underlying water dynamics. This study presents high-resolution along-track observations collected by the autonomous surface vehicle HALOBATES in the German Bight. The measurements capture co-located horizontal density gradients in both the ocean skin layer, sampled at approximately 80 μm, and the near-surface layer (NSL) at 1 m depth. The results reveal partial decoupling and significant attenuation of horizontal density gradients between the skin and the NSL (slope = 0.41 ± 0.008, R² = 0.18), alongside a persistent baseline of skin-layer spatial variability that produces a higher front-detection rate than in the NSL (8.3 % against 7.2 %). Density fronts at the skin are disproportionately salinity-driven rather than temperature-driven (44.1 % salinity-only and 33.7 % temperature-only), more so than at the NSL (39.9 % against 31.8 %). Tidal phase significantly modulates gradient intensity at both layers. Random Forest machine learning coupled with SHapley Additive exPlanations identified skin temperature and salinity as the dominant predictors of gradient intensity, while atmospheric pressure, wind speed, and air-sea temperature difference contributed substantially. Nevertheless, predictive skill across the full range of gradient intensities remained limited (test R² = 0.19). These findings demonstrate that submesoscale density gradients at the skin layer can decouple substantially from underlying upper-ocean structures. This vertical disconnect highlights potential biases in satellite validation and proxy-based approaches that rely on sea-surface skin measurements to infer submesoscale dynamics in complex coastal waters.