ESA SNAP toolbox? It should be possible to make and share a graph [describing an algorithm] that does all the necessary operations.
Apparently it drops to 8-bit while still requiring far too much RAM and processing time (see #1631 above) for 3-channel color. Note there seems to be no way of sharing 16-bit products through the forum -- upload a 16-bit png as image (or attachment) and it will display (and download) as 8-bit.
Attachments are restricted to gif, jpg, mpg, pdf, png, and txt. Tif is not supported even as 8-bit. Our monitors can only display three channels of 8-bit. Even a colorless grayscale is displayed that way, equal values of R,G, and B.
A graph could only take things so far. Each image requires manual contrast adjustment because of variable illumination, cloud cover, and region of interest (ROI). For example, optimizing to ice stream features (ie velocity tracking) sacrifices image quality of rocks without separate masking with an alpha channel prior to processing.
S2A is not dedicated or optimized to ice sheets by any means. One minute it is over Jakobshavn, the next over the Amazon. The ESA takes care of a whole lot of issues but necessarily leaves some to the end user.
The issue is, of the information initially gathered by the satellite's sensors,
how much are you willing to throw away? Some of that information is explicit (readily visible) and some implicit (present, but has to be drawn out).
If all you want to do is monitor calving front position, you can do that with the crude Landsat or DMI preview images. There is no need to download anything, just take a screenshot in your web browser.
If you want to measure slight seasonal or year-on-year acceleration of the Jakobshavn Isbrae (ie is trouble brewing?), you need the full resolution of S2A and an optimal channel (which we previously determined was band 4; it gets into the shadows better). Color is great for melt but contributes adversely to matching displaced and distorted surface features.
Because the imagery is really expensive and enhancement is really cheap, it makes sense to intervene slightly between applications of successive algorithms. Here you'll get a sub-optimal outcome applying automatic pattern recognition to a sub-optimal image.
Does it matter? Yes because
it is all about the error bars. We're trying to push out the best possible real-time prediction here; the uncertainty is an integral part of that.
The image below shows S2A_R111_V20160624T152911.22WEB.B04.tiff right out of the box. Note it is only using a fraction of the 16-bit range available to it. However dumbing it down to 8-bit makes matters considerably worse (histogram on right). Never round off your precision at the first step, save that until the very end.