Accurate sea surface temperature (SST) forecasting is key for understanding marine and climatic dynamics, but remains challenging in high-variability regions such as coastal zones. Deep learning techniques have recently surpassed traditional numerical methods in computational efficiency and accuracy in prediction tasks. In particular, graph neural networks (GNNs) have demonstrated outstanding performance in forecasting climate variables and are attracting interest for modeling ocean dynamics. This work aims to adapt a GNN, originally designed for atmospheric data, to predict the temperature at the ocean surface. However, this type of neural network typically relies on regular meshes, which struggle to capture nonlinear oceanographic processes. Therefore, we propose to use a physically-informed mesh that adapts node density based on the bathymetry of the sea, prioritizing coastal areas. Our method integrates satellite-derived SST data with flexible graph topologies by restructuring latent representations through physics-aware graphs. The model is optimized with the L4 SST satellite images dataset from the Copernicus Marine Service. The results demonstrate that adaptive meshes reduce forecasting errors compared to regular grids, particularly near the coast. This approach bridges geospatial data and graph-based learning, showing that node allocation based on static forcings enhances model performance. The results highlight the potential of geometric deep learning for operational oceanography, offering improved interpretability and accuracy in complex geophysical systems based on remote sensing.