Compaction of Global Data Fields
Abstract
Two methods of compacting global data fields are studied, both individually and in combination, and errors associated with the methods are systematically examined. The first scheme consists of expanding the data into an empirical orthogonal function (EOF) series in the vertical, then truncating the series at a selected number of terms. Because the EOFs are ordered by decreasing variance explained, this reduces the number of degrees of freedom while retaining most of the important vertical structure information. The second technique used is bit reduction, in which appropriately scaled data (here, spectral coefficients) are converted to integer form, with the scaling factor chosen so that the maximum data value is the largest integer expressible by some desired number of bits. Examination of compaction errors for various EOF truncations and bit scalings indicates that one important result of bit reduction is to set to zero all coefficients with magnitude below a certain threshold, causing EOF truncation up to a given point to have no impact on errors. Based on a somewhat arbitrarily selected maximum allowable RMS temperature error of 1 deg C, a compaction factor of approximately two is obtainable from EOF truncation alone, and an additional factor of three from bit reduction (32 bits to 10), assuming half precision words. Future work is needed, however, to determine the generality of these results.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1990
- Accession Number
- ADA225856
Entities
People
- A. H. Van Tuyl
Organizations
- United States Naval Research Laboratory