Fill NaN cell with mean of eight surrounding cells in grid data in matlab

I am trying to fill the NaN values of grid data with 8 surrounding values but could not understand, what is going wrong here in my matlab code. The data is 752*891*11 %%% (11 years precipitation of 752*891 cells).

for i = 2 : 752 
    for j = 2 : 891 
        for k = 1 : 11
            if isnan(data(i,j,k)) == 1
                data(i,j,k) = nanmean(nanmean(data(i-2:i+2,j-2:j+2,k-2:k+2)));
            end
        end
    end
end

Thanks in advance for help

Answer 1

[Jan]

If you want to replace a scalar based on a 3D array, you either need 3 nanmean calls:

nanmean(nanmean(nanmean(data(i-1:i+1, j-1:j+1, k-1:k+1))))

with i-1:i+1, instead of i-2:i+2. This would be easier:

if isnan(data(i,j,k))   % "== 1" is not needed
  tmp         = data(i-1:i+1, j-1:j+1, k-1:k+1);  % 3D block
  data(i,j,k) = nanmean(tmp(:));                  % Make it a vector
end

But this is a 3x3x3 neighborhood with 27 elements, not 8. I assume you mean:

if isnan(data(i,j,k))   % "== 1" is not needed
  tmp1        = data(i-1:i+1, j, k);
  tmp2        = data(i:i, j-1:j+1, k);
  tmp3        = data(i, j, k-1:k+1);
  tmp         = [tmp1(:); tmp2(:); tmp3(:)];  % Create a vector
  data(i,j,k) = nanmean(tmp);
end

Because the center point data(i,j,k) is included 3 times, but ignored by nanmean, you have 6 neighbors now, not 8.

So please explain again, what you exactly want.

[Shakir Hussain]

Seems you got my point, i,j,k (where i and j are lat and lon of certain grid and k is year, e.g lat(12.02),lon(105.27), year(2001). Yes, we have 6 neighbors and if we replace 1 with 2 then we will have 10 neighbors(i.e i-2:i+2, j-2:j+2 value of k). In other words we will fill the NaN cell of 2001 data with mean of 10 neighbors, am i right? Note: We want to replace the NaN with the mean of neighbors (i,j) of a certain year K.

One more thing beside this query is, few months of the certain year are missing and I want to fill that gap with monthly mean of respective year, so how could i do it?

Than you in advance

[Jan]

10 neighbors? I do not know a neighborhood model with 10 elements. I'm not sure what you are asking for.

[Shakir Hussain]

Dear Jan, I mean to say mean of 12, 10,9 and 5 neighbors (preference starts from 12 and last option is 5) in this way to fill the NaN of certain year. Kindly also mention how is work at edges of data.

It is work for single year for i = 2 : 752 for j = 2 : 891 if isnan(sampledata(i,j)) == 1 sampledata(i,j) = nanmean(nanmean(SRTM(i-1:i+1,j-1:j+1))); end end end But, I have to take its loop for 11 years and I could not figure out it at this stage?

I tried your suggested three means way (nanmean(nanmean(nanmean(data(i-1:i+1, j-1:j+1, k-1:k+1))))) but it pop up with an error (Index exceeds matrix dimensions)

[Jan]

I still do not get, which neighborhood you want. Now you mention 4 different ones and some preferences, but no criterion to decide for one. Using nanmean three times creates a mean over a 3D block, but you show 2D blocks. I do not understand, where the 11 years come into play. Maybe all you want is:

for i = 2 : 752 
    for j = 2 : 891 
        for k = 1 : 11
            if isnan(data(i,j,k))
                tmp         = data(i-2:i+2, j-2:j+2, k);
                data(i,j,k) = nanmean(tmp);
            end
        end
    end
end

Such that there is no averaging between the years.

[Shakir Hussain]

The above four preference is for single year and wants to repeat the same way for 11 years. We have both positive and negative values of temperature and your above given code is pop up with (Subscript indices must either be real positive integers or logicals.) error.

[Jan]

When you access the index i-2:i+2, the loop for i must start at 3, not at 2, because 0 is not a valid index. And the loop must stop at size(data, 1) - 2. Equivalently for j.

Please post a copy of the complete message, if you mention an error in the forum.

[Shakir Hussain]

Thank you jan The problem has solved

Answer 2

[Walter Roberson]

means = conv2(YourMatrix, [1 1 1;1 0 1;1 1 1]/8,'same') ;
mask = isnan(YourMatrix);
YourMatrix(mask) = means(mask);

No loops needed.

Note: you would need a little adjustment to handle a nan on the edge of the matrix, to calculate the means properly.

[Jan]

This does not remove the NaNs, because conv2 replies NaN when any element of the neighborhood is NaN. Setting the NaNs to 0 at first does not work correctly, if several elements in the neighborhood are NaN.

[Walter Roberson]

Ah, good point, the 0 * nan at the point itself makes the conv2() result into nan.

[Jan]

A solution to use conv2 with NaNs:

nanX    = isnan(X);
X(nanX) = 0;
mask    = [1 1 1; 1 0 1; 1 1 1];
means   = conv2(X,     mask, 'same') ./ ...
          conv2(~nanX, mask, 'same');
X(nanX) = means(nanX);

But I'm not sure, how this can be applied to the OP's problem. A neighborhood of 6 or 10(?) elements is wanted.

[Image Analyst]

Just change the mask shape and elements to be whatever is wanted.