.. _example_16:

(16) Gridding of data, continued
--------------------------------

:doc:`contour </contour>` (for contouring) and
:doc:`triangulate </triangulate>` (for gridding) use the
simplest method of interpolating data: a Delaunay triangulation (see
Example :ref:`example_12`) which forms *z(x, y)* as a
union of planar triangular facets. One advantage of this method is that
it will not extrapolate *z(x, y)* beyond the convex hull of the
input (*x, y*) data. Another is that it will not estimate a *z* value
above or below the local bounds on any triangle. A disadvantage is that
the *z(x, y)* surface is not differentiable, but has sharp kinks
at triangle edges and thus also along contours. This may not look
physically reasonable, but it can be filtered later (last panel below).
:doc:`surface </surface>` can be used to generate a
higher-order (smooth and differentiable) interpolation of
*z(x, y)* onto a grid, after which the grid may be illustrated
(:doc:`grdcontour </grdcontour>`, :doc:`grdimage </grdimage>`,
:doc:`grdview </grdview>`).
:doc:`surface </surface>` will interpolate to all (*x,
y*) points in a rectangular region, and thus will extrapolate beyond the
convex hull of the data. However, this can be masked out in various ways
(see Example :ref:`example_15`).

A more serious objection is that :doc:`surface </surface>` may estimate *z* values
outside the local range of the data (note area near *x* = 0.8, *y* =
5.3). This commonly happens when the default tension value of zero is
used to create a "minimum curvature" (most smooth) interpolant.
:doc:`surface </surface>` can be used with non-zero
tension to partially overcome this problem. The limiting value
*tension = 1* should approximate the triangulation, while a value
between 0 and 1 may yield a good compromise between the above two cases.
A value of 0.5 is shown in the Figure :ref:`fig_ex16`. A side
effect of the tension is that it tends to make the contours turn near
the edges of the domain so that they approach the edge from a
perpendicular direction. A solution is to use
:doc:`surface </surface>` in a larger area and then use
:doc:`grdcut </grdcut>` to cut out the desired smaller
area. Another way to achieve a compromise is to interpolate the data to
a grid and then filter the grid using :doc:`grdfft </grdfft>` or
:doc:`grdfilter </grdfilter>`. The latter can handle
grids containing "NaN" values and it can do median and mode filters as
well as convolutions. Shown here is :doc:`triangulate </triangulate>` followed by
:doc:`grdfilter </grdfilter>`. Note that the filter has
done some extrapolation beyond the convex hull of the original *x, y*
values. The "best" smooth approximation of *z(x, y)* depends on
the errors in the data and the physical laws obeyed by *z*. GMT cannot
always do the "best" thing but it offers great flexibility through its
combinations of tools. We illustrate all four solutions using a CPT
that contains color fills, predefined patterns for interval (900,925)
and NaN, an image pattern for interval (875,900), and a "skip slice"
request for interval (700,725).  Again, we use :doc:`subplot </subplot>` to set up
and place the four panels, and place the colorbar beneath the subplot.

.. literalinclude:: /_verbatim/ex16.txt
   :language: bash

.. _fig_ex16:

.. figure:: /_images/ex16.*
   :width: 500 px
   :align: center

   More ways to grid data
