naturalneighbor 0.2.1

Dependencies

• Python 2.7 or 3.4+

• Numpy (has been tested with 1.13+)

Demonstration

Natural neighbor interpolation can be more accurate than linear barycentric interpolation (Scipy’s default) for smoothly varying functions.

Also, the final result looks better.

Note that the natural neighbor values usually are extrapolated; they were cut off in the demo to fairly compare with Scipy’s linear barycentric method, which does not extrapolate.

Usage

This module exposes a single function, griddata.

The API for naturalneighbor.griddata is similar to scipy.interpolate.griddata. Unlike Scipy, the third argument is not a dense mgrid, but instead is just the ranges that would have been passed to mgrid. This is because the discrete Sibson approach requires the interpolated points to lie on an evenly spaced grid.

import scipy.interpolate
import numpy as np

import naturalneighbor

num_points = 10
num_dimensions = 3
points = np.random.rand(num_points, num_dimensions)
values = np.random.rand(num_points)

grids = tuple(np.mgrid[0:100:1, 0:50:100j, 0:100:2])
scipy_interpolated_values = scipy.interpolate.griddata(points, values, grids)

grid_ranges = [[0, 100, 1], [0, 50, 100j], [0, 100, 2]]
nn_interpolated_values = naturalneighbor.griddata(points, values, grid_ranges)

Future Work

• Provide options for extrapolation handling

• Support floats and complex numbers (only support doubles at the moment)

• Support 2D (only support 3D)

• Add documentation with discussion on limitations of discrete sibson’s method

• Uncomment cpplint from tox.ini and cleanup C++ code

• Generalize the threading model (currently it uses 8 threads—one for each quadrant)