Monitoring variability in coastal upwelling calls for tools that are physically consistent and operationally robust. We build a physics-based reference index from alongshore wind stress and use deep learning to predict this upwelling index directly from atmospheric drivers. Concretely, a compact spatiotemporal ConvLSTM with a lightweight readout learns the mapping from sea-level pressure, near-surface winds, and Coriolis scaling to a monthly coastal upwelling indicator. The model reproduces held-out variability with high fidelity and yields coastal patterns consistent with expected ocean responses, such as surface cooling and enhanced chlorophyll-a during stronger upwelling. To enhance interpretability and regional utility, we further distill a parsimonious coastal equation via Elastic Net that links atmospheric forcing to the predicted index. The resulting framework couples mechanistic grounding with deep learning skill, delivering an interpretable, regionally tuned indicator for the Moroccan and Iberian coasts, and providing a practical basis for monitoring, interseasonal comparison, and scenario-oriented analyses.