I have an 8-bit raster (TIFF, IMG, and FGDB rasters all exhibit the same behavior) with the pixel type of unsigned integer. In ArcMap 10.2, the value field is being incorrectly identified as a double, so it can't be used in tools like Tabulate Area that require an integer field for the zone field. In ArcMap 10.3 and 10.4.1, the value field is correctly identified as a long. The problem seems to be an introduced one because the value field in the original raster is correctly identified as a long. The original raster was the 2011 NLCD Land Cover (Multi-Resolution Land Characteristics Consortium (MRLC)), which was then clipped using the Clip tool (Clip—Help | ArcGIS for Desktop), and the result of this is where the value field is misidentified as a double.
In summary:
Has anyone experienced this bug, or can you reproduce it?
Over time Esri has been improving the results of saving rasters native to the FGDB, so I'm not surprised to see that later versions are working better. Bit depth is always a challenge when copying rasters between formats.
Grid format is inherently either 32-bit integer or 32-bit float, and it is the original format used by the raster engine that underlies all this stuff, so often processing through grid format is a good workaround - write to esri grid and then run Copy Raster to your target format with the appropriate pixel_type specified.
You also may have better luck using Copy Raster instead of the interactive Data > Export option in ArcMap, specifying the pixel_type on the output.