Focal / Moving Window statistics - open source

Question

I am currently running a process that involves the creation of many "Focal Statistics" rasters in ArcGIS. These rasters use varying-radius circular search areas and calculate three different statistics: Mean, Sum, and Standard Deviation (depending on the input data source). They also must be created in a programmatic manner - either from Python, R, or Bash/Command Line.

I would like to move this to an open source alternative but am having trouble finding an open source equivalent to ArcGIS's "Focal Statistics". The raster::focal() function in R is terribly slow, and for the size of my data it will not work. SAGA GIS has the 'simple filter' tool, but that will only calculate the Mean, not the other two statistics.

Does anyone have an alternative that replicates the functionality of "Focal Statistics" in a reasonably fast, free, and programmatic manner with a circular search window?

Answer 1

Within GRASS GIS you can use r.neighbors which looks at each cell in a raster input file, and examines the values assigned to the cells in some user-defined "neighborhood" around it. It outputs a new raster map layer in which each cell is assigned a value that is some (user-specified) function of the values in that cell's neighborhood. Using the -c flag you can run a circular filter.

To call the functionality from outside, the grass-session Python package is interesting (usage example).

Answer 2

For staying in R, take a look at the velox package. For performance, it has moved processing into C+, using Rcpp and the Boost Geometry libraries. Operations such as extract, focal, aggregation and rasterization benchmark notably faster than comparable raster package functions.

Unfortunately, it does not look like there is a focal standard deviation per se but, you could submit a development request through the GitHub page or, write a Rcpp function yourself and add it to the package.

Answer 3

terra::focal() is much faster than raster::focal(). But in some cases, still too slow. Another great tool for moving window calculations: pkfilter from pktools. From the docs:

This utility implements spatial and spectral filtering for raster data. In the spatial domain (X, Y), the filter typically involves a rectangular convolution kernel (moving window). To avoid image shifting, the size of the window should be odd (3, 5, 7, ...). You can set the window sizes in X and Y directions separately with the options -dx and -dy. A circular kernel (disc) is applied if option -circ is set.

For example, this creates a circular moving window of 11 x 11 cells calculating the mean per cell:

pkfilter -i inputfile.tif -o output.tif -dx 11 -dy 11 -f mean -circ