Data collected by research vessels, satellites, and surface drifters have been mapped to uniform spatial grids with resolution appropriate to the original data density. This section presents the results of the data mapping projects which provided data products for the physical regionalisation.
As part of the Division of Oceanography's Oceans-EEZ Analysis System, an archive of vertical profiles of temperature, salinity, oxygen and nutrients has been collated comprising over 130,000 stations in the seas adjacent to Australia. The data has been interpolated to 38 standard depths and flagged according to data checks undertaken by CSIRO and the National Oceanographic Data Centre (NODC). This standard depth hydrographic dataset represents the basic database available to oceanographers to determine spatial and temporal distribution of physical oceanographic conditions.
Sea surface temperature (SST) data from the NOAA/NASA Pathfinder Program were used to compute a high-resolution climatology of surface temperature mean and variability. The Pathfinder project data processed by Smith et al. (1996) are from a reanalysis of the global area coverage (GAC) data obtained by the AVHRR radiometer aboard the NOAA polar orbiter satellites. The reanalysis employs new algorithms for radiometer calibration and cloud detection. The data are daily "best SST" estimates on a 9 km resolution grid, obtained after the application of cloud detection tests using visible and infra-red (IR) channel thresholds, spatial coherence thresholds, and comparison to a reference.
Determining large scale patterns of physical properties that might delineate distinctive regions of the Australian EEZ is most directly achieved by reference to regularly gridded maps of oceanic properties. The irregular sampling density and accuracy of ocean observations, and lack of statistical stationarity, generally makes the production of maps a difficult task. To map the 130,000 surface temperature observations available in the Australian EEZ to a fine grid resolution of 25km (over 20,000 grid points) incurs a large computational cost. A variety of algorithms for estimating gridded values from irregular data have been described in the statistical and oceanographic literature. These include binned averages, weighted averages, local least squares polynomial fitting, natural or smoothing splines, and Gauss-Markov estimates.
The Gauss-Markov estimation method is popular in oceanographic applications where it is generally referred to as optimal interpolation or objective mapping. The Gauss-Markov estimate at a given point is a weighted sum of surrounding data where the weights are determined from a priori specified noise and signal covariance. It is optimal in the sense that it gives the estimate with the lowest mean squared error of all possible linear estimates, assuming the a priori signal and noise variance is valid. Disadvantages are that the method is computationally intensive for large datasets, and estimation of the data covariance may be difficult.
An alternative approach, which we employ here, is to apply the weighted least squares quadratic smoother of Cleveland and Devlin (1988), known as a "loess" smoother. The computational demands of the method are less that those for Gauss-Markov estimation, and Chelton and Schlax (1994) show that the filtering characteristics of the loess smoother are nearly as good as Gauss-Markov. In addition, the formulation of the loess smoother allows flexibility in the polynomial model to which the data is fitted. Here, in addition to fitting a quadratic function of the horizontal coordinates to data for each standard depth in the hydrographic dataset, we also apply a fit to a simple harmonic time dependence (annual and semi-annual cycles) according to the time of year the observations were made (but neglecting interannual variability). This step has the advantage of reducing potential bias in the estimated mean associated with a tendency for more ocean observations to be made during summer.
The loess smoother quadratic fit is weighted according to a distance metric r defining the separation of observations and estimation (grid) point. We use the horizontal distance between points for r and apply no weight in time. The quadratic fit is then weighted according to (1-(r/R)3)3 where R defines the maximum radius of points to include in the fit. Following Cleveland and Devlin (1988) we choose an R that varies with grid point so that a predetermined number of points (usually 400 though this is varied in some situations of unusual data density) is used for every estimate. This has the attractive property of altering the effective bandwidth of the loess smoother to match data density. In nearshore regions where ship observations are more frequent, the loess smoother automatically decreases the area from which data is taken for the estimate. This feature enables us to produce gridded estimates of the highest possible spatial resolution, though it does have the disadvantage of clouding the actual smoothing scale.
An appendix to this report contains all the ocean property maps constructed for use in the physical regionalisation analysis, as well as maps of the two properties which were computed but not used in the analysis. Maps of the distribution of source data are also provided to give an indication of mapping accuracy. Note that where data is sparse, very long length scales will be used so that mapping may draw on data from outside the region shown.
The resources demanded in making ocean observations from research vessels or ships of opportunity make it impossible to determine mean subsurface ocean conditions at high spatial or temporal resolution. However, observations from satellite-based instruments make it possible to determine some features of the surface ocean climatology at much higher resolution that from in situ observations. Notable among these is sea surface temperature, which is routinely observed by radiometers on the NOAA polar orbiting meteorological satellites. The AVHRR instrument only observes SST in cloud free regions.
We have computed the mean SST, and variability at annual and semi-annual periods, of 4 years (1987-1990) of daily "best SST" data from the NOAA/NASA Pathfinder Program. The fitting of time harmonics to the data time series minimises a sampling bias in the data resulting from increased cloudiness during winter. This climatology is produced at the full resolution of the Pathfinder data; nominally 9km spacing at mid-latitudes.
Further analysis of the daily data has been completed as part of the Division of Oceanography's Oceans-EEZ Analysis System. By applying optimal interpolation techniques (Gauss-Markov theory) to the time series data, we have produced 10-day optimal averages without cloud gaps. This time resolution appears to be the shortest possible interval over which meaningful "composites" can be determined from the Pathfinder data.
The satellite SST data offer the highest resolution of any physical oceanographic dataset that covers the entire EEZ. Many 10-day optimal average maps show mesoscale structures with length scales of order 100 km. However, at the annual and semiannual time scales considered here these length scales do not appear, and the satellite SST data are barely distinguishable from maps computed from in-situ hydrographic observations. This gives us confidence that the mapping procedure employed for the hydrographic data has achieved a spatial resolution appropriate to the time scales of interest. For this reason, the satellite SST data have not been included in the regionalisation analysis.
Next Chapter: 10. Biological Regionalisation: Datasets