Filter image that contains NaNs in Matlab?

Question

I have a 2d array (doubles) representing some data, and it has a bunch of NaNs in it. The contour plot of the data looks like this:

Answer 1

Want to improve this post? Add citations from reputable sources by editing the post. Posts with unsourced content may be edited or deleted.

nanfilter does exactly the same thing with nanconv when filtering as long as the filter is the same. If you get the nan values before you use nanfilter and then add the back to the after-filtered matrix, you will get the same result with what you get from nanconv with the option 'nanout', as long as you use the same filter.

Answer 2

Okay without using your plot function, I can still give you a solution. What you want to do is find all the new NaN's and replace it with the original unfiltered data (assuming it is correct). While it's not filtered, it's better than reducing the domain of your contour image.

% Toy Example Data
rfVals= rand(100,100);
rfVals(1:2,:) = nan;
rfVals(:,1:2) = nan;

% Create and Apply Filter
filtWidth = 3;
imageFilter=fspecial('gaussian',filtWidth,filtWidth);
dataFiltered = imfilter(rfVals,imageFilter,'symmetric','conv');
sum(sum(isnan( dataFiltered ) ) )

% Replace New NaN with Unfiltered Data
newnan = ~isnan( rfVals) & isnan( dataFiltered );
dataFiltered( newnan ) = rfVals( newnan );
sum(sum(isnan( rfVals) ) )
sum(sum(isnan( dataFiltered ) ) )

Detect new NaN using the following code. You can also probably use the xor function.

newnan = ~isnan( rfVals) & isnan( dataFiltered );

Then this line sets the indices in dataFiltered to the values in rfVals

dataFiltered( newnan ) = rfVals( newnan );

Results

From the lines printed in the console and my code, you can see that the number of NaN in dataFiltered is reduced from 688 to 396 as was the number of NaN in rfVals.

ans =
   688
ans =
   396
ans =
   396

Alternate Solution 1

You can also use a smaller filter near the edges by specifying a smaller kernel and merging it after, but if you just want valid data with minimal code, my main solution will work.

Alternate Solution 2

An alternate approach is to pad/replace the NaN values with zero or some constant you want so that it will work, then truncate it. However for signal processing/filtering, you will probably want my main solution.

Answer 3

One approach would be to replace the NaN values with nearest-neighbor interpolates using scatteredInterpolant (or TriScatteredInterp in older MATLAB versions) before performing the filtering, then replacing those points again with NaN values afterward. This would be akin to filtering a full 2-D array using the 'replicate' argument as opposed to the 'symmetric' argument as a boundary option for imfilter (i.e. you're replicating as opposed to reflecting values at the jagged NaN boundary).

Here's what the code would look like:

% Make your filter:
filtWidth = 7;
imageFilter = fspecial('gaussian', filtWidth, filtWidth);

% Interpolate new values for Nans:
nanMask = isnan(rfVals);
[r, c] = find(~nanMask);
[rNan, cNan] = find(nanMask);
F = scatteredInterpolant(c, r, rfVals(~nanMask), 'nearest');
interpVals = F(cNan, rNan);
data = rfVals;
data(nanMask) = interpVals;

% Filter the data, replacing Nans afterward:
dataFiltered = imfilter(data, imageFilter, 'replicate', 'conv');
dataFiltered(nanMask) = nan;

Answer 4

The technique I ended up using was the function nanconv.m at Matlab's File Exchange. It does exactly what I was looking for: it runs the filter in a way that ignores the NaNs just the way that Matlab's built-in function nanmean does. This is a hard to decipher from the documentation of the function, which is a tad cryptic.

Here's how I use it:

filtWidth = 7;
filtSigma = 5;
imageFilter=fspecial('gaussian',filtWidth,filtSigma);
dataFiltered = nanconv(data,imageFilter, 'nanout');

I'm pasting the nanconv function below (it is covered by the BSD license). I will post images etc when I get a chance, just wanted to post what I ended up doing for anyone curious about what I did.

Comparison to other answers

Using gnovice's solution the results look intuitively very nice, but there are some quantitative blips on the edges that were a concern. In practice, the extrapolation of the image beyond the edges led to many spuriously high values at the edges of my data.

Using krisdestruction's suggestion of replacing the missing bits with the original data, also looks pretty decent (especially for very small filters), but (by design) you end up with unfiltered data at the edges, which is a problem for my application.

nanconv

function c = nanconv(a, k, varargin)
% NANCONV Convolution in 1D or 2D ignoring NaNs.
%   C = NANCONV(A, K) convolves A and K, correcting for any NaN values
%   in the input vector A. The result is the same size as A (as though you
%   called 'conv' or 'conv2' with the 'same' shape).
%
%   C = NANCONV(A, K, 'param1', 'param2', ...) specifies one or more of the following:
%     'edge'     - Apply edge correction to the output.
%     'noedge'   - Do not apply edge correction to the output (default).
%     'nanout'   - The result C should have NaNs in the same places as A.
%     'nonanout' - The result C should have ignored NaNs removed (default).
%                  Even with this option, C will have NaN values where the
%                  number of consecutive NaNs is too large to ignore.
%     '2d'       - Treat the input vectors as 2D matrices (default).
%     '1d'       - Treat the input vectors as 1D vectors.
%                  This option only matters if 'a' or 'k' is a row vector,
%                  and the other is a column vector. Otherwise, this
%                  option has no effect.
%
%   NANCONV works by running 'conv2' either two or three times. The first
%   time is run on the original input signals A and K, except all the
%   NaN values in A are replaced with zeros. The 'same' input argument is
%   used so the output is the same size as A. The second convolution is
%   done between a matrix the same size as A, except with zeros wherever
%   there is a NaN value in A, and ones everywhere else. The output from
%   the first convolution is normalized by the output from the second 
%   convolution. This corrects for missing (NaN) values in A, but it has
%   the side effect of correcting for edge effects due to the assumption of
%   zero padding during convolution. When the optional 'noedge' parameter
%   is included, the convolution is run a third time, this time on a matrix
%   of all ones the same size as A. The output from this third convolution
%   is used to restore the edge effects. The 'noedge' parameter is enabled
%   by default so that the output from 'nanconv' is identical to the output
%   from 'conv2' when the input argument A has no NaN values.
%
% See also conv, conv2
%
% AUTHOR: Benjamin Kraus (bkraus@bu.edu, ben@benkraus.com)
% Copyright (c) 2013, Benjamin Kraus
% $Id: nanconv.m 4861 2013-05-27 03:16:22Z bkraus $

% Process input arguments
for arg = 1:nargin-2
    switch lower(varargin{arg})
        case 'edge'; edge = true; % Apply edge correction
        case 'noedge'; edge = false; % Do not apply edge correction
        case {'same','full','valid'}; shape = varargin{arg}; % Specify shape
        case 'nanout'; nanout = true; % Include original NaNs in the output.
        case 'nonanout'; nanout = false; % Do not include NaNs in the output.
        case {'2d','is2d'}; is1D = false; % Treat the input as 2D
        case {'1d','is1d'}; is1D = true; % Treat the input as 1D
    end
end

% Apply default options when necessary.
if(exist('edge','var')~=1); edge = false; end
if(exist('nanout','var')~=1); nanout = false; end
if(exist('is1D','var')~=1); is1D = false; end
if(exist('shape','var')~=1); shape = 'same';
elseif(~strcmp(shape,'same'))
    error([mfilename ':NotImplemented'],'Shape ''%s'' not implemented',shape);
end

% Get the size of 'a' for use later.
sza = size(a);

% If 1D, then convert them both to columns.
% This modification only matters if 'a' or 'k' is a row vector, and the
% other is a column vector. Otherwise, this argument has no effect.
if(is1D);
    if(~isvector(a) || ~isvector(k))
        error('MATLAB:conv:AorBNotVector','A and B must be vectors.');
    end
    a = a(:); k = k(:);
end

% Flat function for comparison.
o = ones(size(a));

% Flat function with NaNs for comparison.
on = ones(size(a));

% Find all the NaNs in the input.
n = isnan(a);

% Replace NaNs with zero, both in 'a' and 'on'.
a(n) = 0;
on(n) = 0;

% Check that the filter does not have NaNs.
if(any(isnan(k)));
    error([mfilename ':NaNinFilter'],'Filter (k) contains NaN values.');
end

% Calculate what a 'flat' function looks like after convolution.
if(any(n(:)) || edge)
    flat = conv2(on,k,shape);
else flat = o;
end

% The line above will automatically include a correction for edge effects,
% so remove that correction if the user does not want it.
if(any(n(:)) && ~edge); flat = flat./conv2(o,k,shape); end

% Do the actual convolution
c = conv2(a,k,shape)./flat;

% If requested, replace output values with NaNs corresponding to input.
if(nanout); c(n) = NaN; end

% If 1D, convert back to the original shape.
if(is1D && sza(1) == 1); c = c.'; end

end