Tuesday, 11 February 2003: 11:00 AM
Classifying El Nino and La Nina events using the Kohonen maps
Neural networks have been used in recent studies to evaluate the predictability of the El Nino/Southern Oscillation (ENSO) phenomenon and to set up operational procedures of ENSO forecasts. The present study aims at using neural networks-more especially the Kohonen maps- to classify El Nino and La Nina events over the past century (1900-2001). We use sea surface temperature anomalies (SSTA) from Kaplan (1998) averaged in the four major regions of the tropical Pacific: Nino1=90W-80W/5S-10S, Nino2=90W-80W/0-5S, Nino3=150W-90W/5N-5S, and Nino4=160E-150W/5N-5S. The four Ni-o indices define the entry layer. Then the topological map of Kohonen is formed of 100 linear neurones (i.e. a 10x10 map) all connected to each neurone of the entry layer. At the first time step of the training period, the weights of all the neurones are initialized randomly. Then an iteration of the training period is defined by two steps : (i) a competition between all the neurones (i.e. each neurone computes its "state" relatively to a vector of observations selected randomly among data, and the best matching unit, bmu, is selected; (ii) an adaptation of all the neurones of the map following a "Temperature Law" i.e. with an adjustment of the weights of each neurone function of the distance on the map to the bmu. Moreover the characteristics of that function evolves at each iteration with a shape sharper around the bmu at each new iteration i.e. with a shorter distance of influence. That methodology allows to distort the initial regular grid of neurones i.e. to cluster the observations. In fact, we used a probablistic version of the topological maps called PRSOM and developped at LODYC. Using PRSOM allows to compute second order statistics and consequently to forecast El Nino and La Nina events according to Bayes theory. The results of the training period shows that the map is organized in coherent regions. Considering the map as a square of 10 by 10 neurones, we found that the upper corner to the left is associated with La Nina events when all the Nino indices are negative, the lower corner to the right (opposite) is associated with El Nino conditions (positive indices). Briefly, each region of the map is associated to a different state of the Nino indices i.e. a displacement on the map can be decomposed as changes in the east-west SST gradient and changes in the central Pacific SSTA. We studied the time trajectories of past El Nino and La Nina events on the Kohonen map allowing us to synthesize the large amount of data distinguishing major trajectories corresponding to different scenarii of ENSO events. More than a powerful tool to visualize the characteritics of growth and decay of ENSO events, the topological maps are also suggested to have a potential for ENSO forecasts.