This letter developes and validates a machine learning approach to forecast sea level anomalies (SLAs) derived from satellite altimetry in the tropical Pacific Ocean. The empirical orthogonal function (EOF), also known as principal component analysis, was used to extract dominant signals and reduce the dimensionality of data sets. Such dimensionality was decreased by describing spatial patterns (EOFs) and the corresponding temporal domains [principal components (PCs)]. Support vector regression (SVR) was employed to predict the time series obtained from the leading PCs. Thereafter, the temporal and spatial SLAs from the proposed EOFs were reconstructed to represent the spatiotemporal SLA prediction. Finally, the prediction result was compared with that of the conventional autoregressive integrated moving average (ARIMA) model. Both models reached satisfactory sea level predictions. Even so, intercomparison of the obtained results showed that the SVR significantly (P = 0.012) outperformed the ARIMA model in sea level forecasting. That is, a considerably low root-mean-square error was attained for the differences between the predicted and observed mean SLAs.
All Science Journal Classification (ASJC) codes