Visualization improvements for datashader examples

[PeterDSteinberg]

Recently much of the elm-examples repo (soon deprecated) were moved to this repo's ./examples directory as part of making elm and related tools easier to get started with. In the examples folder there is a LANDSAT notebook that uses datashader visualization. Today, @jbednar mentioned that the LANDSAT datashader work so far could be improved with more attention to color scales in RGB image plots of normalized band differences or other computed rasters. We should modify that notebook in the following ways:

• [ ] Better color scales for objective interpretation of the data. Currently the interpretation in part depends on which band is chosen to be R, G, B, etc. See also these links on color scale mapping:
  • http://peterkovesi.com/projects/colourmaps/index.html
  • https://arxiv.org/abs/1509.03700
• [ ] More documentation on what is being done in the plots, such as legends, more meaningful titles, units, etc. Specifically here are two parts that need docs:
  • [ ] The choice to downsample, currently set at a resolution of (600, 600), needs to be explained as a quick decimation for purposes of a demo with limitations due to spectral averaging the original ca (7500, 7500) pixel bands.
  • [ ] The last output plot for hierarchical prediction needs an explanation to the effect of "this is the final prediction created by a binary matrix encoding of all of the ensemble members' K-Means output" and it could also explain that hierarchical modeling and other multi-model techniques will be developed further in Phase II (the notebook is a prototype idea for now on those ideas).
• [ ] Interactive widgets that help check the classifier. This could optionally be deferred until we begin NASA UI work.
• [ ] Check the calculation of NDVI in the notebook and similar normalized differences. I think I noticed that at least NDVI ratio is using the wrong band combination (see also this description of LANDSAT 8 bands - bottom of page)

[PeterDSteinberg]

Eventually we can add cross links between datashader and elm examples

[jbednar]

Oops; I just added a similar issue on the EarthIO site: https://github.com/ContinuumIO/earthio/issues/14

I'll close that one, but what it said was:

As described in section 6 (p31) of https://arxiv.org/pdf/1509.03700.pdf, the rgb color space is useful for constructing true-color images, but it is a very misleading way to make false-color images from three matrices that do not correspond to the R, G, and B wavelength bands, like those for most Landsat data, because the perceptual salience of these three different wavelengths of light is very different. As a result, the appearance of a three-band image can differ greatly depending on whether each band is mapped to R, G, or B. That paper outlines a better way to make a ternary image, I believe with accompanying matlab code at peterkovesi.com, and we should probably provide this code for Python to promote good practice. The colorcet package already contains a lot of Peter's work, and would probably be a good place to put this code too, and then maybe a simple way could be provided to make a HoloViews object (hv.TernaryImage?).