These scripts are modified versions of the core functionality in the dsmart package for R. All props to Nathan Odgers who originated the concept and coded it in python, and Brendan Malone who ported it to R . If you want the original python code, hit @nathanodgers up on twitter. If you want the official R package, contact Brendan at brendan.malone@sydney.edu.au. That code is under a GPL-2 license, which this material inherits.
These modifications started with a desire for area-proportional sampling within polygons rather than a flat rate, and then I couldn't leave well enough alone. Along the way I've managed to improve the main function's stability, if not its speed, and I've made the outputs a little easier to work with. As I write this, its happily chugging through 6000 polygons and a 14-raster covariate stack with about 1.4M cells per layer.
- area-proportional sampling, with a minimum rate that still allows minor soil classes a look-in
- improved memory management
- all inputs start in sp or raster objects and stay there
- outputs handle NA values correctly e.g. masked out coastal areas
- option of outputs just as rds objects