How are NaN/Inf values interpreted when calculating statistics in ENVI?

Calculating statistics on images that contain NaN (not-a-number) or Inf (infinity) values can produce different results depending on whether or not a mask is used in the calculation. This Help Article describes why you might get different results when calculating statistics if you do or do not use a mask band, and when it might be necessary to use a mask.

In ENVI, NaN (not-a-number) and Inf (infinity) pixels in an image are ignored when calculating statistics via the /NAN keyword to TOTAL(), MIN(), and MAX(). However, the pixel count is still for the whole image. ENVI does not explicitly calculate how many NaN/Inf pixels there are in the image. Instead it will ignore any that are there. For example,

mean = total(data) / n_elements(data)

which ignores the NaN/Inf pixel values when calculating the total, but the number of pixels used as the divisor includes all pixels (including the pixels containing NaN/Inf values).

However, when you use a mask band to mask out the pixels containing NaN/Inf values, the number of pixels used in the calculation is exactly the number of pixels that are good in the mask (that are turned 'on' or have a value of 1 in the mask). The NaN/Inf pixels are then not included in the calculation, so the difference is essentially:

mean = total(data[ptr]) / n_elements(ptr)

where ptr is the pointer reference that contains the actual number of 'good' pixels to include in the statistics calculation.

If you have a lot of pixels containing NaN/Inf values in your image, then a mask should be used to obtain accurate statistics. If only a few pixels contain NaN/Inf values, the statistics will not be substantially influenced and a mask might not be necessary.

You can build a NaN/Inf mask using ENVI's Masking tool under Basic Tools -> Build Mask.

If your image contains NaN/Inf values, it is likely a mask will be necessary for most other processing in ENVI to obtain accurate results.

Review on 12/31/2013 MM