There are a fistful of ways to specify metadata when using things like NetCDF, CF conventions in a NetCDF File, NcML, THREDDS and ncISO, which can lead to confusion. As it stands now, there is substantial overlap in the possible semantics that can be expressed in different places in the technology stack and no clear precedence model (or even a naming scheme/controlled naming conventions for the different metadata tags). Folks are working on it at Unidata, NOAA, USGS, and other places, but this is not a done deal. This wiki page lists some of the more common methods, discussing what the end-user will see if one method overrides another within the stack of technology.
The folks at NOAA Geo-IDE have done a lot of work and thinking (hopefully not in that order) on this type of issue, with a particular eye towards automating the translation of metadata from one standard of expression to another (such as from NetCDF CF to ISO 19115-2). See their wiki article on the subject for more. There's a figure at the bottom of this page that's particularly relevant. Note that the figure references CF (Climate and Forecast conventions) at the bottom of its precedence sequence. CF is a set of conventions for specifying metadata in general and is used in the context of their wiki to refer to the metadata within the NetCDF file (I think).
An important benefit of clearly establishing a precedence model for metadata is to enable the relation of one format/model of metadata to another. Unidata has made recommendations for defining metadata that allow its tools (and those of others in this field) to automate translation into other metadata formats, which greatly enhances discoverability of the data content. The metadata tags in these recommendations are drawn from those that are part of their own NetCDF conventions and from the CF conventions. Other metadata formats that these tools can translate into include
A somewhat old and incomplete table shows correspondance between metadata tags from different standards.
probably need to get into ncISO/THREDDSISO/UDDC and all that here. not sure if that should be a separate page.
The NOAA Geo-IDE wiki points out that the CF conventions are optimized for data use (access and processing). The metadata, although relatively sparse, is sufficient to automatically derive numerous other pieces of metadata relatively easily. See this section for an example. The upshot is that while many of us are using CF as a great starting point for describing our data content, we might step back and think about augmenting that metadata to help improve discoverability of our data content.
In addtion, one can specify metadata content in any of a number of places, some better than others depending on the context. It would probably be good to find or develop some guidance on best practices about this for our data producers. NOAA Geo-IDE is probably a good place to start. There are a number of large data production projects (UAF, IOOS) that have aligned themselves with the operational principles developed within Geo-IDE. Unidata also provides guidance (PowerPoint, web site) on best practices for NetCDF in particular.
By developing metadata that can at ultimately be expressed according to ISO 19115, we ensure that any CSW-compliant client software can discover and access our data services (see this subsection of the Geo-IDE wiki . This is in addition to ensuring to say nothing of all the client software that can deal with data servers that converse according to ISO 19115, or any of the other metadata formats that one can cross-walk to from there.