thinking about data handling and analysis. This is followed by discussions of some issues relating to specific classes of data, and a summary of areas to which the statistics community may be well equipped to contribute.

The Committee on Data Management And Computing define five general classes of spacecraft data, based on the degree of processing involved (CODMAC, 1982, and subsequent refinements):

• *	Level 0 -The raw data stream from the spacecraft, as received at Earth

• *	Level 1 -Measured radiances, geometrically and radiometrically calibrated

• *	Level 2 -Geophysical parameters, at the highest resolution available

• *	Level 3 -Averaged data, providing spatially and temporally "uniform" coverage

• *	Level 4 -Data produced by a theoretical model, possibly with measurements as inputs

This paper focuses on Level 2 and Level 3 data, which are the main concerns of most global change research scientists working on EOS instrument teams. Level 2 products are reported on an orbit-by-orbit basis. For a polar-orbiting satellite such as EOS, the Level 2 sampling of Earth is highly non-uniform in space and time, with coverage at high latitudes much more frequent than near the equator. Level 2 data are needed when accuracy at high spatial resolution is more important than uniformity of coverage. These situations arise routinely for validation studies of the satellite observations, in the analysis of field campaign data, and when addressing other local-and regional-scale problems with satellite data.

The spatially and temporally uniform Level 3 data are needed for global-scale budget calculations, and for any problem that involves deriving new quantities from two or more measurements which have different sampling characteristics. To derive a Level 3 product from Level 2 data, spatial and temporal scales must be chosen. It is to this issue that we turn next.

The creation of Level 3 data has traditionally involved the selection of a global, 2- or 3-dimensional spatial grid, possibly a time interval as well, and "binning" the Level 2 data into the grid cells. The binning process for large data sets usually entails taking the arithmetic mean and standard deviation of all Level 2 data points failing into a grid cell, with possible trimming of outliers or of measurements flagged as "low quality" for other reasons. Typically, all points included in a grid cell average are given equal weight. Occasionally a median value will be used in place of the mean.

The leading contender for the standard EOS Level 3 grid is a rectangular-based scheme similar to one that has been used by the Earth Radiation Budget Experiment (ERBE) (Green and Wielicki, 1995a). In the proposed implementation for EOS, the Earth is divided zonally into 1.25 degree strips (about 140 km in width). Each strip is then divided into an integral number of quadrilaterals, each approximately 140 km in length, with the origin at the Greenwich meridian. This produces a nearly equal-area grid.