Coastal upwelling prediction with a mixture of neural networks

For the analysis and prediction of coastal upwelling in a study region off the northwest African coast, artificial neural networks are applied to wind data and remotely sensed sea surface temperature (SST) data. The coastal upwelling phenomenon is seen as an input-output system, built by the dependence between local wind events as input and upwelling, quantified by the SST index as output. To gain a priori knowledge about the input-output system, it is studied from a theoretical point of view with coastal upwelling simulations, which are performed with a three-dimensional (3D) ocean circulation model on artificial wind data. The dynamics observed in these studies are found to be bounded and dependent on previous states as well as nonlinear, which motivates the application of artificial neural network (ANN) techniques as a nonlinear modeling technique. To cope with this type of system and with the nonstationarity of the parameters used, a mixture of neural networks is applied for system analysis and prediction. The output space of the system is segmented in an unsupervised manner, while the modular neural networks are specializing on one segment each. It is shown that different dynamic structures in the input-output system are detected by the mixture approach and prediction is improved. The synergetic approach presented, using numerical models to gain a priori knowledge for the choice and design of a neural prediction tool, is appealing for implementing simulations or modeling into an operational application, for instance, geographical information systems (GISs) or decision support systems (DSSs).