In this paper, the predictive performance of a standard Recurrent Neural Network (RNN) applied to sea surface temperature data is evaluated. The novelty of this work is a neighborhood-based scheme for feeding spatiotemporal information into the RNN model. Our approach is compared against another RNN trained on purely temporal sequences and on an average baseline. Results demonstrate that incorporating spatial neighborhoods improves predictive performance, highlighting the relevance of this strategy for forecasting oceanographic variables from reanalysis datasets
