|
Percentage of Haze in an
image: The method needs hazy and clear areas
within the image! If only limited clear areas are
available (image has predominantly haze and
clouds) the algorithm fails (or might
over-correct certain areas).
|
|
Type of ‘Clouds’: In
the scientific paper* on which the algorithm
is based the authors speak about "an image
transform to characterize and compensate
for spatial variations in thin cloud contamination...".
Accordingly only thin clouds/haze can be reduced
in its appearance [a good example for
"haze" within an image can be seen here].
An a priori test if a hazy
image area ("cloud") can be haze
corrected is to check if the RED and NIR-bands
(e.g. for IKONOS = bands 3,4) are transparent
(the ground can be seen). If YES, it is haze (in
the sense of ATCOR = thin cloud contamination)
which ATCOR can 'tackle' if NO, it is an impermeable
cloud and can not be haze corrected.
To test the above: display
Bands 2, 3 and 4 without applying any stretch. If
the background becomes visible everything is set
for an ATCOR success.
Examples:
|
|
|
|
|
|
|
|
|
|
|
|
In both images on the left (displayed in the
2xStDev Stretch) the cloud seems to have the same
density. If displayed in band combination 4,3,2 as
R,G,B without any histogram stretch it becomes
obvious that in the example on the bottom ATCOR
can not reduce the cloud cover much besides in a
small transition zone.
|
|
|
If there is no good correlation
between the
blue and red band (r
< 0.8) the method also fails to produce good
results
Info
on the correlation coefficient 'r' [External
Link].
|
|
The method can also be used for data without the blue band but will yield
less perfect
results
by using the green band as a substitute. This is true for
sensors
like e.g. IRS.
|
|
The
algorithm is
not suitable for haze over water.
|
Originally the Haze Correction is an optional pre-processing step to the
Athmospheric Correction only. While it can be run independently without the
Atmospheric Correction, it has limitations (which the Atmospheric Correction does not have): Size constraint: an image can not have more than 8K * 8K in
size.
|
|
*Zhang, Y.,
Guindon, B., and Cihlar, J., "An image
transform to characterize and compensate for spatial
variations in thin cloud contamination of Landsat
images", Remote Sensing of Environment, Vol.
82, 173-187 (2002).
|
