set NaN as another color than default using imagesc

Hello, I have a matrix filled probability numbers (i.e. ranging from 0 to 1) or NaN when the probability is not computed. I would like to display this matrix as a color table (e.g. using imagesc), in order to have a quick visualisation of the result. The colorbar range is thus set as 0 to 1 since I am interested in probability values. However, I would like the NaN fields to appear with another color than the default "-inf" (here the color of 0 since the down limit for the color is set for value 0), for example gray. How can I do this? Thank you very much, Gaelle

Answer 1

[Kelly Kearney]

Do you need to use imagesc, or are you open to using pcolor? The latter will leave NaNs blank, so they will appear the same color as the axis background.

Example:

 data = rand(10);
 data(data > 0.9) = NaN;
 [nr,nc] = size(data);

 subplot(2,1,1); 
 imagesc(data); 
 subplot(2,1,2); 
 pcolor([data nan(nr,1); nan(1,nc+1)]);
 shading flat;
 set(gca, 'ydir', 'reverse');

(The NaN-padding in the pcolor call is so the last row and column are shown, similar to imagesc)

[SeanC]

Hi Kelly,

This solution was helpful thanks. I was wondering if there is a way I can manipulate the scale of the plot using this code - i.e. the equivalent of the clims argument in imagesc. This is because I have different plots but want them all to take the same scale.

Thanks,

Sean

[Kelly Kearney]

Just set the 'clim' property of the axes:

data = rand(10);
data(data > 0.9) = NaN;
[nr,nc] = size(data);

ax(1) = subplot(2,1,1); 
imagesc(data); 
ax(2) = subplot(2,1,2); 
pcolor([data nan(nr,1); nan(1,nc+1)]);
shading flat;
set(gca, 'ydir', 'reverse');

set(ax, 'clim', [0 1]);

[SeanC]

Thanks!

Answer 2

[Charles Krzysik]

If you want to use imagesc rather than pcolor, you can set the AlphaData property to zero everywhere you have a NaN. This will show the background wherever there is no data; you can then additionally set the background colour, if you so choose. In this example I'm setting it to black:

imAlpha=ones(size(Data_Array));
imAlpha(isnan(Data_Array))=0;
imagesc(Data_Array,'AlphaData',imAlpha);
set(gca,'color',0*[1 1 1]);

[JoeB]

This is a really nice answer. Reading in netCDF files with missing values yields NaNs in MATLAB, so this allows me to make quick maps with the NaNs masked out.

[teng zhang]

It is so nice for your answer.

[TheStranger]

This is the most useful comment of this thread, yet not with the most votes!

[Maria Cristina Araya Rodriguez]

I agree, many thanks, in fact this is a good solution when using a ".cpt" colour scale.

Answer 3

[Lauren]

imagesc(data,'AlphaData',~isnan(data))

[Huayan Wang]

I have to say this is a genius solution!

Thank you so much! This helps me a lot!

[Xiaowei Zhou]

Excellent idea.

[Walter Roberson]

you might need double(~isnan(data)) as alphadata cannot be logical .

[Shakir Hussain]

How to use it for array data, it is giving error for array data(3d)

Error using image
Bad property value found.
Object Name: image
Property Name: 'AlphaData'.

Error in imagesc (line 20)
    hh = image(varargin{:},'CDataMapping','scaled');

[Walter Roberson]

please show your code calling imagesc

[jichong han]

perfect

[zhou weiyan]

perfect

[Walter Roberson]

If you have RGB data, some elements of which might be NaN, then

%with sufficiently new versions of MATLAB
scaled_data = rescale(YourRGBData, 0, 1);
alpha_data = 0 + ~any(isnan(YourRGBData), 3);
image(scaled_data, 'AlphaData', alpha_data);

%older versions of MATLAB
scaled_data = mat2gray(YourRGBData);
alpha_data = 0 + ~any(isnan(YourRGBData), 3);
image(scaled_data, 'AlphaData', alpha_data);

In practice if your RGB data is single or double precision data that uses only integral values 0 to 255, but also has some NaN, then you would probably use slightly different code,

scaled_data = uint8(YourRGBData);
alpha_data = 0 + ~any(isnan(YourRGBData), 3);
image(scaled_data, 'AlphaData', alpha_data);

Answer 4

[Evan]

You can use the isnan function to find the indices of all NaNs, then set those elements to another value:

Example:

>>A = [4 NaN 3;NaN 2 1];
>>A(isnan(A)) = 255

A =
     4   255     3
   255     2     1

[Gaëlle]

Hi, Thank you for your answer, but this will not work in my case, because I am setting the limit of the colorbar (using caxis([0,1]) ). If I set, as you proposed, the NaN fields as e.g. 255, the color assigned to them will be equivalent to the color assigned to the value 1 (here the upper bound of the colorbar). I would rather code something that assign a specific color (grey) in the case of a NaN field. Is there any way to do this? Gaelle

[Evan]

It depends. You always run the risk of setting the pixels in question to a value that is already used if you don't know anything specific about your image data.

Are you just wanting to visualize where your data returns as NaN? If so, I would do one of two things: I would plot my "good" data as grayscale, then turn the NaN pixels to a easily noticeable color like red. Alternatively, if you must have or prefer your data in color, you could plot your image then overlay a scatter plot where your elements are NaN.

Here's an example showing the first way:

A = rand(50); %random data
for i = 1:10 %turn some random points into NaNs
   A(randi(50),randi(50),:) = NaN;
end
R = A(:,:,1); % turn your data into a "pseudo-gray" rgb image.
G = A(:,:,1); 
B = A(:,:,1);
R(isnan(A(:,:,1))) = 1; %turn NaNs red
G(isnan(A(:,:,1))) = 0;
B(isnan(A(:,:,1))) = 0;
A(:,:,1) = R; %combine color slices into A
A(:,:,2) = G;
A(:,:,3) = B;
imagesc(A)

Here's an example for the second:

A = rand(50); %random data
for i = 1:10 %turn some random points into NaNs
   A(randi(50),randi(50),:) = NaN;
end
imagesc(A)
hold on
[ii jj] = find(isnan(A));
scatter(ii,jj,'ok','MarkerEdgeColor',[1 1 1],'MarkerFaceColor',[0 0 0],'LineWidth',2,'SizeData',100)

[Gaëlle]

Hi, No, the aim is still to plot the probability values. For more clarity: this matrix M is a square matrix, where each element (i,j) corresponds to a comparison (a probability calculus) between an element i and j, e.g. the element M(1,2) gives the probability value of the comparison between element 1 and 2 (note that in mx case M(1,2)≠M(2,1) ).

As for the NaN, they are here simply because I want to keep the size of the matrix similar across all calculation and display. They appear because some comparison do not arise. For example, if the jth element does not exist, the entire line j and row j will be set as NaN. For example, if I want to compare 4 elements, but the 2nd one does not exist yet, the matrix will look like:

if true
  % code
0.3 NaN 0.4 0.1 
NaN NaN NaN NaN 
0.2 NaN 0.2 0.5 
0.5 NaN 0.6 0.4
end

I thus want to display this matrix, but setting the NaN values as something specific (e.g. a grey value so that it resembles the color of the figure background). Is this more clear? Gaelle

[Gaëlle]

What you showed me is almost that. I just would have liked to keep the 'jet' map instead of turning every value into a grey scale. But I should manage to find a trick. Thank you, Gaelle

[Gaëlle]

But one thing which is bad using your way: the colorbar scale is not set according to the values of the probabilities. So actually, I cannot use your trick...

[Adil Masood]

Thanks you for a smart solution. There is just one correction: Instead of

scatter(ii,jj,'ok','MarkerEdgeColor',[1 1 1],'MarkerFaceColor',[0 0 0],'LineWidth',2,'SizeData',100)

it should be

scatter(jj,ii,'ok','MarkerEdgeColor',[1 1 1],'MarkerFaceColor',[0 0 0],'LineWidth',2,'SizeData',100)

Its because imagesc() plots elements of matrix indexed as (row,column), while scatter() handles elements as (x,y).

Answer 5

[Gaëlle]

Hi, I am open to use pcolor, it just seemed more complicated to use than imagesc. I will try to implement it. Thank you!

Answer 6

[Ben]

If using pcolor, NaNs are set to the axis background color. So you can set your colormap to color non-NaNs, and set the axis background color to set the color for NaNs. For example, to set NaNs black you can use: set(gca,'color','k')

[Walter Roberson]

For pcolor the nan are not exactly set to the axes background color: instead a hole is created that allows whatever is under to be visible. If nothing else is there that would be the axes background, but there could also be a different graphics object there.