To appear
Previous: 9 Compressing a TIN
We have studied many compression algorithms for regularly gridded terrain elevation files, including both generic image processing methods, and some semicustom ones. The generic image processing compression algorithms perform so well that there appears no need to design algorithms specifically for elevation data. This also allows us to take advantage of the continuing progress in image processing algorithms. For example, the best algorithms that we studied are all quite new. With sp_compress, for example, USGS DEMs compress down to an average of 2 bpp. There is a wide variation in the compressability of different data; however the relative performance of various methods on any particular file does not vary widely.
Compression should not hinder interactive use of the data; partitioning one test file into 256 blocks before compressing increased thae total size by only 13%.
Several open questions remain. Can these methods be improved by fine-tuning internal parameters for the statistical properties of elevation data, if those properties are, in fact, different than the scenes that the algorithms were designed for? Can the very low bit-rates at which progdecd reports an exact reconstruction be extended into a new compression program at those rates? Since lossy compression is much more compact, often far below 1 bpp at moderate mean squared errors, how much lossiness can we tolerate before essential properties of the data such as drainage patterns and visibility are damaged?
\