Non-Gaussian Process Modeling of Argo Float Data

Jonas Wallin, Lund University, Department of statistics, Lund, Sweden
Argo floats measure seawater temperature in the upper 2000 m of the ocean. The statistical analysis of the data is challenging due to several reasons: non-stationary, the space-time-depth dimensions, and also the large size of data. Recent research in interpolation of Argo Float data have been using Gaussian space-time processes to construct temperature maps at varying depth. In said research it was noted that the Gaussian processes does not mimic the behaviour of the data, since the tail of distribution is heavier than that of the Normal distribution. Additionally it was shown that using student-t measurement error did not fully correct the issue, hence indicating that the seawater temperature is a non-Gaussian process. Recently we have developed multivariate non-Gaussian processes. This allows for using non-Gaussian processes and jointly modeling different depth of the ocean temperature. The idea is that observations of different depths will give more information about each other thus contributing to a better representation of each field. Early analysis indicates that both the multivariate model and the non-Gaussian process improves the fit of the data of the univariate Gaussian process. The same multivariate models should also be possible to apply when jointly modelling seawater temperature and salinity (which also Argo measures).

Further we are currently exploring non-Gaussian space time models which we intend to apply to the Argo data, as previous analysis has shown that space-time modeling gives large improvement compared to pure spatial modeling, but the data still is non-Gaussian.