mean value of subarrays

I have a 3-dimensinal array ( 1483 rows, 417 columns and 4 bands). I am trying to get a mean value for each sub-array of 92rows*92columns. I have some zero values that I want to exclude when the calcualtion of mean is being done. I used the following code but it's not correct.my data type is uint32. i apperciate your help.

Answer 1

[Sean de Wolski]

A couple of calls to convn. The first call will determine the number of zeros; the second will calculate the sum of all values in the sub array; array divide the second by the first.

If you don't want to do it to EVERY POSSIBLE subarray, use blkproc or blockproc with function handle

 @(x)sum(x(:))./(sum(logical(x(:)))

Also be aware that since your data type is uint32, if the sum of the 92*92*4 values exceeds 4,294,967,295 it will truncate to 4,294,967,295 and your mean will be wrong.

[Paulo Silva]

Sean have you noticed the amount of related questions that come up lately? you answered some of them :)

[Hassan]

thanks Sean for the help. I typed your command and repalced it with my data but it seems it's not the way to use it. can you explain more please?

to have an integer number of sub-arrays (16*5) I removed 11 rows and added 43 columns (with zero values). I used the following code but it's not working.

new_array=zeros(16*5,4)

kk=1

for i=1:4

for j=1:1472

for k=1:460

n_zeros=sum(Data(j:j+91,k:k+91,i)==0)

sum_=sum(Data(j:j+91,k:k+91,i))

mean_=sum_/n_zeros

new_array(kk)=mean_

if (j<rows && k<cols)

j=j+92;

k=k+92;

kk=kk+1;

end

end

end

end

[Hassan]

about sub-arrays(16*5), I meant a new array with 16 rows and 5 columns. I should have said sub-arrays(92*92).

[Sean de Wolski]

No reason for FOR-Loops. What do you mean by "not the way to use it"?

You could always do the for-loops but it's a lot of extra work for what is very simply; two convolutions.

Why are you dividing by the number of zeros? Shouldn't it be the number of non-zeros?

DON'T overwrite the built-in function MEAN!!!!

[Sean de Wolski]

End of the semester Paulo ;)

[Hassan]

About using Loops, unfortunately I don't know how to use other functions to avoid using Loops.

I don't know the meaning of comman you wrote (@(x)sum(x(:))./(sum(logical(x(:)))).

I replaced 'x' by 'Data' (which is the name of my data) but it didnt give me anything.

Sorry it should be number of non-zeros (n_zeros=sum(new_array(j:j+91,k:k+91,i)~=0)).

still, it's not working. I'm getting the folling error:

??? Index exceeds matrix dimensions.

Error in ==> upscaling at 90

n_zeros=sum(new_array(j:j+91,k:k+91,i)~=0)

[Sean de Wolski]

doc blkproc. The function I wrote would be the handle you would feed it.

Something like:

mean_wout_zeros = blkproc(Data,[92 92],@(x)sum(x(:))./(sum(logical(x(:))));

But I just remembered that blkproc is only for 2-dimensions so forget the whole thing.

Answer 2

[Sean de Wolski]

Here's a 2-dimensional example that demonstrates the logic and kernels.

%sample data
A = uint32(repmat(magic(10),[4,1]));
A(A<10) = 0;

%We will use a window size 6x10
window = ones(6,10);
nnzeros = conv2(double(logical(A)),window,'valid'); %you can change this shape do whatever it is you want at the edges.
winsum  = conv2(double(A),window,'valid');
mean_wout_zeros = winsum./nnzeros;

%check that it worked (for first valid value)
isequal(sum(sum(A(1:6,:)))/nnz(A(1:6,:)),mean_wout_zeros(1))

You could convert mean_wout_zeros back to uint32 if you want...

[Hassan]

I think I misexplained the subarray thing. I want a mean value for each 92*92 subarrays which is a new array with a 16 rows and 5 columns. a mean value for the first (1:92,1:92) then the next mean value for the next subarray (1:92,93:184), ...

using the follwing code I get mean value as the same number of array's elements.

A(1472*460 uint32)

%to calculate a mean for a window of 92*92

window = ones(92,92);

[Sean de Wolski]

Where does the 16x5 come from? It seems to be appearing out of the air. For the skipping every 92; do EXACTLY what I have above (with a ones(92) window and extract every 92nd value. I.e. mean_wout_zeros = mean_wout_zeros(1:92:end);

[Hassan]

I am working with a remote sensing image(1483*417*4) which each pixel is 5 meters * 5 meters. i want to average it up for each 92*92 pixels. the new array which has the mean values has 1483/92=16.1 and 417/92=4.5 dimensions. I removed the last 11 rows and added 43 columns (zero values) to create a 16*5 array.

[Hassan]

As I said I need an array of ean values (16 rows,5coulmns,4 abnds) mean values while using the qhat you said above, I got vecotr of 5540 elements.

it seems blkproc,as you wrote the code above, does what I need.

thanks for the help Sean.