While there exists a number of Python functions that can be used to upsample geo-spatial images and grids to a higher spatial resolution, there is a generally a lack in methods that perform grid cell aggregation for downsampling grids to a coarser spatial resolution. Most of the existing aggregation methods assume an integer factor between a source and its target grid so that contributions of partly overlapping grid cell areas don't exist.

The gridtools.resampling module provides the high-performance regridding functions resample_2d(), upsample_2d(), and downsample_2d().

Downsampling can take into account partial contributions of source grid cells for a given target grid cell; it performs a weighted aggregation of grid cell contributions.

• Method DS_FIRST: Take first valid source grid cell, ignore contribution areas.
• Method DS_LAST: Take last valid source grid cell, ignore contribution areas.
• Method DS_MEAN: Compute average of all valid source grid cells, with weights given by contribution area.
• Method DS_MODE: Compute most frequently seen valid source grid cell, with frequency given by contribution area. Note that this method can use an additional keyword argument mode_rank which can be used to generate the "n-th Mode". See downsample_2d().
• Method DS_VAR: Compute the biased weighted estimator of variance (see https://en.wikipedia.org/wiki/Mean_square_weighted_deviation), with weights given by contribution area.
• Method DS_STD: Compute the corresponding standard deviation to the biased weighted estimator of variance which is basically the square root of the result of method DS_VAR.

The methods DS_MEAN, DS_VAR DS_STD are most useful for downsampling grids whose cell values represent continuous values, e.g. temperatures, radiation.

The methods DS_FIRST, DS_LAST DS_MODE are most useful for downsampling grids whose cell values represent classes, e.g. surface types, flags.

Currently, only two upsampling methods are provided:

• Method US_NEAREST: Take nearest source grid cell, even if it is invalid.
• Method US_LINEAR: Bi-linear interpolation between the 4 nearest source grid cells.

The module provides functions that allow filling grid cells whose values are not finite by means of the numpy.isfinite() function. Two gap-filling methods are available:

• Function fillgaps_lowpass_2d(): Fills cell values by averaging values of direct neighbours using a given kernel. This is repeated for the whole grid until all gaps are filled.
  • Pros: Simple and obviously working well for mostly isolated, single cell gaps.
  • Cons: Naive. Relatively slow, if gaps form larger connected areas. In this case gap border patterns propagate into gap area centers at multiples of 45 degree angles, producing strange visual artifacts, and usually an implausible distribution of filled values.
• Method fillgaps_multiscale_2d(): Similar to fillgaps_lowpass_2d() but tries to get around its disadvantages: Cell values are filled in by averaging values of direct neighbours. Then the resulting grid is downsampled by a factor of two. If the downsampled grid still has gaps, the procedure is repeated recursively until the downsampled version contains no more gaps or it comprises only a single cell. Each gap-filled, downsampled grid serves as a source for gaps in the upsampled, 2x higher resolution grid until the original resolution is reached and all gaps are filled (or none).

• All resampling methods assume the target grids to be in the same coordinate space, hence only a grid scaling is applied where the geometric boundaries and coverage of source and target remain the same.
• All methods are currently 2D only because our primary goal is to perform spatial resampling.
• Upsampling is currently limited to only two methods. Use existing alternatives instead such as scipy.misc.imresize or similar.

gridtools uses Numba to JIT-compile the resampling functions as it speeds up computations by up to a several hundred times (!) compared to plain Python.

gridtools is tested with Numba in Miniconda or Anaconda environments.

To disable JIT compilation (e.g. for unit-level testing), set environment variable NUMBA_DISABLE_JIT to a non-zero value.

There is an issue in Numba that currently limits its use in certain cases when grids are represented by numpy masked arrays, see numba/numba#1834

From 0.3 to 0.4

• Changed license from GPL to MIT (#1)