Since the different methods display different properties of the data set, the most useful approach is probably to use several of them together. An especially useful combination seems to be first to reduce the amount of data either by clustering or by the SOM, and then to display the reference vectors with some distance-preserving projection method to gain additional insight. In his original paper Sammon suggested that clustering could be used as a front-end to his mapping algorithm [Sammon, Jr., 1969]. If the SOM is used to perform the clustering, there will be two different views to the same data available, which would certainly be useful.
It has also been proposed [Demartines, 1994, Demartines and Hérault, 1997] that in a combination of some vector quantization algorithm followed by metric MDS (or Sammon's mapping), the cost function of the latter could be modified slightly by introducing a decreasing weighting function F:
The function F forces the mapping to concentrate on local distances.