Prediction of long lead monthly three-dimensional ocean temperature using time series gridded Argo data and a deep learning method

Abstract

Ocean temperature is a vital physical variable of the oceans. Accurately predicting the long lead dynamics of the three-dimensional ocean temperature (3D-OT) can help us identify in advance potential extreme events (e.g., droughts and floods) that may be caused by the changes of the 3D-OT, which however remains a challenge. To achieve this goal, a deep learning (DL) model was proposed to make predictions of the monthly 3D-OT for one year ahead using time series gridded Argo data. The DL model is comprised of a one-dimensional convolution (Conv1D) layer which is used for extracting latent features from the time series ocean temperature data, two long short-term memory (LSTM) layers which are used for capturing the long-term temporal dependencies hidden in the 3D-OT based on the features extracted by the Conv1D layer, and a fully-connected layer to output the predictions. The proposed DL model can well model the temporal dependencies and dynamic patterns of the ocean temperature at different spatial locations and in different depths by learning from simply the historical time series gridded Argo data. Experiments conducted in a sub-area of the South Pacific Ocean that predict the monthly 3D-OT with the lead time from 1 to 12 months show that the developed DL model surpasses the persistence model, the AdaBoost model, and the feedforward backpropagation neural network model (BPNN) when compared from multiple spatiotemporal perspectives using multiple statistics, indicating that the proposed DL model is a highly strong model for long lead monthly 3D-OT predictions.

Keywords

• One-dimensional convolution (Conv1D)
• Long short-term memory (LSTM)
• Time series gridded Argo data
• Three-dimensional ocean temperature (3D-OT) prediction