In r-gris/torpor: Lazy Topology

Summary

Automatic reprojection is absolutely fine, but it's not always going to work properly in a number of different ways. These aren't mutually exclusive, but come down to either the CRS metadata or the actual topology of the line segments that link the raw coordinates, their arrangement and distribution and relation to the space in which they are operated on.

1) the data has no projection set (NA), not much we can do (and we shouldn't guess, it might be in the long-lat range but we have no idea of the datum, and it might be in different units anyway) 2) the data has its CRS set, but it's wrong (nothing we can do, it's not possible to auto-check this)

Both of those scenarios are the responsibility of the user, and it both case they need to to find the right CRS and set it.

3) The data has its CRS set, but it's not on WGS84. In this case reprojection either needs the toWGS84 params, or the extra installed datum files for GDAL to do datum transformation.

4) The data has coordinates that are out of bounds in the target, e.g. the dummy seam at the south pole for Antarctica in longlat, this coordinate/line-segment is undefined in Mercator and degenerate in many projections (two identical coordinates with a zero-length segment).

5) The data is cut along a seam or not cut along a seam, i.e. the dateline. These are the wrapping problems, but they can manifest in several different ways.

6) Insufficient curvature along segments to carry over to another projection (great circles, parallels, lines of constant bearing, ... ) generally this is about what does "straight line" mean in this CRS, and what does it mean to transform it to another CRS. In ggplot2 this dealt with in non-geo forms with "coord_munch" that segmentizes line segments (adds redundant vertices so the curve is carried through when transformed). Mike bostock flawed example ...

7) Raster reprojection is slightly different ...

Relation to TopoJSON and rmapshaper, other packages that auto project, e.g. raster, oce, (there are many examples).

Finally, all of this can be dealt with from a "finite element perspective" ... but it's for the future and impacts more broadly than the "should we reproject" issue.

The danger in 2) is that we might reproject their data but the source was completely wrong, so we make it worse.

The danger in 3) is that the toWGS84 params are missing, or incorrect, or that a reprojection operation simply assumed it was WGS84. We get some inaccuracies in the locations and poor provenance on that situation for downstream uses.

r-gris/torpor documentation built on May 26, 2019, 1:34 p.m.