optimize correlation length scale for horizontal dimensions

[JianghuiDu]

Regarding 2D lat-lon analysis. Two questions:

1. Is it possible to optimize the correlation length scales for both latitude and longitude? I know how to optimize len = (len_lat,len_lon)*scale using bestfactorl given byDIVAnd_cv. But I'd like to have len = (len_lat*scale_lat, len_lon*scale_lon) and optimize both scale_lat and scale_lon.

2. It appears when using DIVAnd_cv the ranges explored are always between 10^-1.5 and 10^1.5 for epsilon2 and 10^-1 and 10^1 for len regardless of nl and ne. So is it possible to expand the search area and what exactly do those two numbers control?

[jmbeckers]

1. We did not implement an independant optimization of Lx and Ly assuming that if you know there is an anisotripy, you should be able to specify it to some extend. Otherwise you are adding one more free parameter to calibrate. If you really want to implement it you would need to adapt DIVAnd_cv (but again with all the limitations inherent to the method)

2. Those numbers just controlling the width in L and epsilon space over which you search an optimum. If you want a larger domain you can either manually increase the values in DIVAnd_cv OR move the search window by hand (providing different input values for the central values of L and epsilon)

[jmbeckers]

For the moment, if you really would like to explore the possibility to calibrate both Lx and Ly is to let DIVAnd_cv optimize around that value and then only change by hand the value of Lx for example increase and redo. If you find a better optimum, you are going into the right anistropy direction. If it worsens, change the value of Lx into a lower value than the original one and retry.