Predicting the evolution of sea surface temperature (SST) is essential for applications in weather forecasting, maritime transport, and fisheries. Traditional ocean forecasting methods rely on physics-based numerical models, which face challenges such as data gaps, assimilation difficulties, and computational inefficiencies. Recent advances in Graph Neural Networks (GNNs) have shown promise in improving prediction accuracy and efficiency. In this work, we adapt a GNN model, initially designed for atmospheric forecasting, to oceanographic applications. We focus on the Canary Islands and the northwest African shore regions characterized by strong mesoscale dynamics. Our approach introduces a spatially masked loss function to address ocean-specific challenges like spatial discontinuities and observational data sparsity. We train our model using the L4 SST satellite images dataset from Copernicus Marine Service and compare its performance with state-of-the-art ConvLSTM-based models. Our results indicate that the adapted GNN model effectively captures mesoscale structures and outperforms ConvLSTM in both computational efficiency and accuracy. These findings suggest that graph-based deep learning approaches can overcome key limitations of current oceanographic models and provide a more flexible and scalable solution for forecasting oceanographic variables from satellite images.