Independent component regression for seasonal climate prediction: an efficient way to improve multimodel ensembles

Abstract

The main goal of this study is to improve the seasonal climate prediction of multimodel ensembles. The conventional principal component regression has been used to build a statistical relation between observations and multimodel ensembles. It predicts future climate values when there are a large number of variables, which is a typical issue in climate research field. However, principal component analysis which is prerequired to perform principal component regression assumes that information of the data should be retained by the second moment. This condition would be stringent to climate data. In this paper, we present a new prediction method that is efficient to adapt to non-Gaussian and high-dimensional data. The proposed method is based on a combination of independent component analysis and regularized regression approach. The main benefits of the proposed method are as follows. (1) It explains a statistical relationship between multimodel ensembles and observations, when either one is not normally distributed; and (2) it is capable of evaluating the contribution of climate models for prediction by selecting some specific models that are appropriate. The superiority of the proposed method is demonstrated by the prediction of future precipitation in boreal summer (June-July-August; JJA) for 20 years (1983-2002) on both global and regional scales.