Downsampling algorithm comparison
- Mark Taylor
When downsampling (mapping from a high resolution grid to a low resolution grid) the main issue is what the mapping procedure does to the features with length scales that can not be resolved on the low resolution grid. With interpolation based algorithms, high frequency modes often alias onto lower frequency modes. With cell integrated type algorithms, high frequencies are averaged over the target cell, mapping to their mean value and removing variance.
The page show results from an extreme case: mapping a high frequency mode (TempestRemap’s T16_32 spherical harmonic test function) from ne120pg2 (~40km) to ne3pg2 (~1600km). The Y16_32 function has 16 oscillations in the mid-latitudes which are well represented on the ne120pg2 grid. On the ne3pg2 grid, we have approximately 24 cells along 45N or 45S, and thus this grid is far too coarse to capture the Y16_32 oscillations.
Grid | Contour Plot |
|---|---|
Reference plot: Y16_32 plotted from the ne120pg2 grid This particular function ranges from 1 to 3. | 
|
bilinear map ne120pg2 → ne3pg2 The wave number 16 mode aliases onto a wave number 4 mode that has little similarity to the original function
| 
|
integrated bilinear map ne120pg2 → ne3pg2 The wave still aliases onto a wave number 4 pattern, but the target grid’s large cells mean that any cell integrated approach will be integrating over much of the original oscillation, resulting in a much weaker wave.
|
|
mono / aave map ne120pg2 → ne3pg2 Nearly identical to integrated bilinear
|
|