Theoretical analysis and empirical results indicate that
only when training data from a certain time or space must be
extended through time and space is atmospheric correction
necessary for image classification and many types of change
detection. Rayleigh and other types of scattering could increase
brightness in the visible band (400-700nm). Atmospheric
absorption is the main factor reducing the brightness values of
pixels in the near- and middle-infrared region (700-2400nm).
The Landsat TM near- and middle-infrared bandwidths were
carefully chosen to minimize the effects of atmospheric
absorption. Therefore, if a single temporal Landsat TM data is
atmospherically corrected, it is likely that the primary effect
will be a simple bias adjustment applied separately to each
band. This action would adjust the minimum and maximum
values of each band downward. Training class means extracted
from the single temporal image would change but the training
class variance-covariance matrices should remain invariant
[20]. Therefore, the actual information content used in the
maximum likelihood classification of the dataset would remain
unchanged. The general principle is that atmospheric correction
is not necessary as long as the training data are extracted from
the image under investigation and are not imported from
another image obtained at another place or time. In this study,
two different temporal imagery are analyzed independently
using a maximum likelihood classification algorithm, and then
the resultant images are mosaicked. So it’s not necessary to
atmospherically correct the single temporal remote sensing
data.