Research ArticleOCEANOGRAPHY

Purely satellite data–driven deep learning forecast of complicated tropical instability waves

See allHide authors and affiliations

Science Advances  15 Jul 2020:
Vol. 6, no. 29, eaba1482
DOI: 10.1126/sciadv.aba1482
  • Fig. 1 Architecture of our DL SST forecast model.

    The model consists of a DNN and a time-independent bias correction map, both of which were determined by the samples during the training period. The model receives SST maps at the previous and current time steps and then outputs the SST map at the future time step. The DNN has four stacked composite layers, each of which receives SST maps at different resolutions and has four cascaded convolutional layers. The bias correction map is added to the DNN output to obtain the forecasted SST map.

  • Fig. 2 Accuracy of DL model during the testing period.

    (A) Temporal trend of RMSE, (B) temporal trend of bias, (C) spatial distribution of RMSE, and (D) spatial distribution of bias. The RMSE and bias temporal trends were calculated sample by sample from the forecasting errors at all grids. The RMSE and bias spatial distributions were calculated grid by grid from the forecasting errors of all samples.

  • Fig. 3 Satellite and DL-forecasted SST maps at three consecutive time steps (12, 17, and 22 September 2010) with an interval of 5 days.

    We make the next-time-step SST forecast using satellite SST maps at the previous 13 and current time steps. For example, the forecast on 12 September 2010 is made using 14 satellite SST maps at a 5-day interval between 4 July ... 2 September and 7 September. (A to C) Satellite SST maps at the three time steps and those forecasted by the DL model (D to F).

  • Fig. 4 TIW westward propagation speed during the testing period.

    The speed is estimated on the basis of the maximum detrended cross-correlation along the equator between the satellite MAs at the current time step and the satellite or forecasted MAs at the next time step (see Methods). (A) Speed at the whole 20° latitude band. The speed estimated from satellite/satellite SST MA pairs is denoted by deep brown dashed curve. The forecasted speed estimated from satellite/DL-forecasted SST MA pairs is denoted by green solid curve. The daily Niño3.4 index data are denoted by orange dotted curve. March 12, 2010 is the first day to be forecasted. (B) Distribution of speeds at 2° latitude bands estimated from satellite/satellite SST MA pairs and (C) distribution of forecasted speeds at 2° latitude bands estimated from satellite/DL-forecasted SST MA pairs. The white blanks represent outliers beyond the normal range of 0 to 100 km/day.

  • Fig. 5 Trends of global RMSE and bias of DL model implemented recursively concerning the number of recursive steps.

    (A) RMSE and (B) bias. In the recursive manner, the model-forecasted SST at a future time step is fed back to the model input to make the SST forecast at the next future time step. The recursive steps from 1 to 30 are 5 to 150 days after the current time step. After 14 recursive steps, there is no satellite SST map at the model input, and all input SST maps are forecasted SST maps. The global RMSE and bias of the DL model at each recursive step were calculated from the forecasting errors of all samples at all grids (green solid curve). The DL model was also compared with a long short-term memory (LSTM) model (cyan solid curve), a multilayer perceptron model (magenta dash-dotted curve), and a site-specific linear regression (SSLR) model (blue dashed curve). The detailed comparisons are given in the Supplementary Materials.

  • Fig. 6 Satellite SST maps and those recursively forecasted by DL model at the three consecutive time steps (27 September and 2 and 7 October 2010) subsequent to the final time step in Fig. 3.

    The final time step in Fig. 3 is the first recursive step here, and the three subsequent time steps are the second to fourth recursive steps. (A to C) Satellite SST maps at the three time steps and those forecasted in the recursive manner by the DL model (D to F).

Supplementary Materials

  • Supplementary Materials

    Purely satellite data–driven deep learning forecast of complicated tropical instability waves

    Gang Zheng, Xiaofeng Li, Rong-Hua Zhang, Bin Liu

    Download Supplement

    The PDF file includes:

    • Supplementary text
    • Figs. S1 to S9
    • Legend for movie S1

    Other Supplementary Material for this manuscript includes the following:

    Files in this Data Supplement:

Stay Connected to Science Advances

Navigate This Article