deepsphere
interpolation of unevenly sampled data
Thank you for your work!
It is not completely clear to me what to do if data is severely unevenly sampled, such as in the climate measurement example.
Do I have to interpolate all the data to one of the discretizations (Healpix, equiangular, etc.)? This interpolation will obviously result in incorrect values in regions that are far from points where the true data has been measured. How to deal with such points during training? Or is this being taken care of by the adjacency matrix?
Another question is about the change of locations where we have data. Suppose that the total number of sites where data is measured remains the same from one data sphere in the training set to the next one, but the locations where the data is measured change. By change I don't mean the rotation of the measurement sites on the sphere, which wouldn't be a problem. Instead I mean movements of individual vertices that would affect the euclidean or angular distance between vertices. Would that be a problem for using your method? Does it require the adjacency matrix to be the same for all the data spheres in the training set?
