Abstract
Kaiman filter theory and autoregressive time series are used to map sea level height anomalies in the tropical Pacific. Our Kalman filters are implemented with a linear state space model consisting of evolution equations for the amplitudes of baroclinic Kelvin and Rossby waves and data from the Pacific tide gauge network. In this study, three versions of the Kalman filter are evaluated through examination of the innovation sequences, that is, the time series of differences between the observations and the model predictions before updating. In a properly tuned Kalman filter, one expects the innovation sequence to be white (uncorrelated, with zero mean). A white innovation sequence can thus be taken as an indication that there is no further information to be extracted from the sequence of observations. This is the basis for the frequent use of whiteness, that is, lack of autocorrelation, in the innovation sequence as a performance diagnostic for the Kalman filter.
Our long-wave model embodies the conceptual basis of current understanding of the large-scale behavior of the tropical ocean. When the Kalman filter was used to assimilate sea level anomaly data, we found the resulting innovation sequence to be temporally correlated, that is, nonwhite and well fitted by an autoregressive process with a lag of one month. A simple modification of the way in which sea level height anomaly is represented in terms of the state vector for comparison to observation results in a slight reduction in the temporal correlation of the innovation sequences and closer fits of the model to the observations, but significant autoregressive structure remains in the innovation sequence. This autoregressive structure represents either a deficiency in the model or some source of inconsistency in the data.
When an explicit first-order autoregressive model of the innovation sequence is incorporated into the filter, the new innovation sequence is white. In an experiment with the modified filter in which some data were held back from the assimilation process, the sequences of residuals at the withheld stations were also white. To our knowledge, this has not been achieved before in an ocean data assimilation scheme with real data. Implications of our results for improved estimates of model error statistics and evaluation of adequacy of models are discussed in detail.