Prevent spurious anomalies triggered by distributions fit from noise
WSIM outputs for September show large areas of combined surplus/deficit in arid regions, such as northwestern Chad/northeastern Niger, and the Iran/Iraq border:
Some of these cases are caused by runoff values that are abnormally high but still negligible, e.g., 6e-17 mm with a distribution location of 1e-18 mm. Discussed w/ @tmparris a couple of methods for resolving this:
- supplying a "resolution" parameter to
wsim_fit
, to be considered in the test for number of unique values. In the case described above, all historical runoff values would be seen as identical, and a distribution would not be produced for the pixel. This would prevent the runoff-based anomalies described above, but requires that the code callingwsim_fit
have some knowledge of a reasonable resolution for each parameter. - modifying
wsim_lsm
, which is aware of the scientific quantities being output and their units, to quantize its outputs to an appropriate resolution, e.g., 0.1mm for runoff. This would also cause all historical runoff values to be seen as identical. - modifying
wsim_lsm
to include a resolution as an attribute for each variable, while still writing variables at full precision. This resolution could then be read bywsim_fit
or other code, where appropriate. It doesn't appear that CF conventions provide a standard way to describe variable accuracy.