Need suggestion -- which interpolation method should I use?

Hi all,

I am working with a relatively sparse matrix with a lot of NaN values. As you can see from the attached picture -

What kind of interpolation scheme do you think will fit the best for this kind of case? Any feedback will be greatly appreciated!

[Torsten]

You mean a scheme that interpolates NaN values ? :-)

What does the matrix represent ? The values z of a function z(x,y) on a specified x-y grid ?

[Ashfaq Ahmed]

Hi! Each of the values represent a temperature. Each of the column is a snapshot of a time

[Walter Roberson]

"snapshot of a time" is not enough information for us, as outside observers, to feel confident that the values can reasonably be treated as continuous in row and column. If, for example, the times (columns) are 12 hours apart, then you would expect a repeated high/low pattern rather than a continuous pattern. And if the rows are weather stations in alphabetical order then the rows might not have anything to do with each other.

Answer 1

[Walter Roberson]

Consider https://www.mathworks.com/matlabcentral/fileexchange/4551-inpaint_nans

However, this depends upon the idea that the data is continuous in 2D, rather than each row or each column being independent.

Answer 2

[Image Analyst]

You might try scatteredInterpolant. Just plug in the values that you do have (not the nan ones) and it will give you an interpolant object that you can predict any value. For example you can get predictions for the missing nan values if you want. See attached demo.