Observational data are captured in real time, usually unique and irreplaceable, like brain images, survey data.
Experimental data are from lab equipment, often reproducible, but can be expensive such as chromatograms, microassays.
Simulation data are generated from test models where model and metadata may be more important than output data from the model, for example economic or climate models.
Derived or compiled data result from processing or combining 'raw' data, often reproducible but expensive, like compiled databases, text mining.
Reference or canonical data are a (static or organic) conglomeration or collection of smaller (peer reviewed) datasets, most probably published and curated, such as gene databanks, crystallographic databases.