Hi
First of all thanks for decades of state of the art softwares released by
unidata, specially python-metpy. I have operationally used it specially for
rbf interpolation as :
result=metpy.interpolate.points.interpolate_to_points(points, values,
grid2, interp_type='rbf',rbf_func='linear',rbf_smooth=0.)
by the way I didn't understand method until I reprogrammed it in fortran.
metpy routine is highly paralelised so uses asmuch as possible from
remote server capacities.
I have the following suggestions:
1- First of all, It would be great if we can some-how disable
parallelization of control the consumption of CPU capacity so that other
WRF runs can start when the program is running.
2- RBF method requires to calculate an inverted matrix(with the shape n*n
and n being the number of observations). The majority of calculation (also
the benefit of RBF method) is based on this step. Finally the inverted
matrix is performed dot product with a matrix of station-grid distances
(or type of rbf function implemented on it) which is straight-forward.
If we want to interpolate multiple variables in the same location, we need
to loop over metpy interpolation routine, but if the number and position of
observations do not change, there would be no need to calculate inverted
matrix. Therefore It would be great if the function output includes
inverted matrix(size n*n) so we are not obliged to recalculate it for ech
variable. Also it would be great if it can be integrated as an accessor to
xarray package . It would be great if this method can be implemented over a
confined neighbourhood around grid points for downscaling purposases too,
even thought it (as the easiest and least complex form of Artificial
Neural Network) is not primarily designed for this purpose.
Thanks again and Best Regards
Amin Fazl Kazemi
Iranian Meteorological Organization/National Center of climate and Crought
risk management (NDC/IRIMO)
www.ndc.irimo.ir
