scattering. This is scattering that occurs at the molecular level and is responsible
for the blue sky when looking upwards from the Earth’s surface. Similarly, when
looking downwards from a sensor through the atmosphere, the apparent reflectance
of the surface tends to be excessively blue. Water vapor and other aerosols also
contribute to absorption and scattering and can be highly variable across sites and
times.
Not all approaches require atmospheric correction, in particular atmospheric
correction methods that affect every pixel in the same way across an image do not
affect the information content of the image. However, time series analysis may
require atmospheric effects be removed to align images radiometrically, and
simple visual interpretation may benefit from the removal of path radiance (haze).
Below we review the different approaches that can be taken and what is involved
in a practical sense.
Empirical line and dark pixel subtraction—Strictly speaking, Eq. 4.1 is an
approximation as it ignores the more complex scattering pathways and spatial
variation in surface reflectance. Nevertheless Eq. 4.1 is the basis of a number of
atmospheric correction schemes. Importantly Eq. 4.1 establishes a linear relationship at each wavelength between the surface reflectance, L w (k), (what we
want) and the at-sensor reflectance, L t (k), (what we have). The two components
that must be deduced in each band are the scaling factor and offset, t(k) and L a (k).
If reference ground truth reflectances are available then these parameters can be
estimated by linear regression for each band and this is the basis of the so-called
‘empirical line’ atmospheric correction. Smith and Milton (1999) review the
requirements to accurately perform this correction, in particular reference targets
must be homogenous and substantially larger than the pixel size to avoid adjacency or point-spread function (PSF) effects (Milovich et al. 1995). In addition
reference targets should encompass the range of reflectances in each band, to avoid
extrapolation. Empirical line is also suitable for high spatial resolution satellite
sensors (Karpouzli and Malthus 2003). A basic form of empirical line is dark pixel
subtraction (Mather 1999), which assumes the darkest band values in the image
represent a surface reflectance of zero. This is not recommended for most processing algorithms that require atmospheric correction (e.g., inversion methods,
Sect. 4.3.5) and will have little or no effect for most classification approaches
(Caplosini et al. 2003).
Cloud shadow method—Lee et al. (2007) present a convenient method that can
be applied if an image contains deep water areas, clouds, and cloud shadows on
deep water. The method also requires an estimation of the ratio of direct to total
irradiance on the surface, which can be obtained from freely available and simple
to use radiative transfer models such as SBDART (Ricchiazzi et al. 1998) or
libRadtran (Mayer and Kylling 2005). The method is relatively insensitive to this
estimation so while it requires specifying atmospheric constituents a ‘standard
atmosphere’ assumption may be sufficient.
Radiative transfer modeling—A more sophisticated approach is to use a
radiative transfer model, either to evaluate t(k) and L a (k) or to parameterize a more
complicated inversion that more fully captures multiple scattering photon paths.
90
J. D. Hedley
for the blue sky when looking upwards from the Earth’s surface. Similarly, when
looking downwards from a sensor through the atmosphere, the apparent reflectance
of the surface tends to be excessively blue. Water vapor and other aerosols also
contribute to absorption and scattering and can be highly variable across sites and
times.
Not all approaches require atmospheric correction, in particular atmospheric
correction methods that affect every pixel in the same way across an image do not
affect the information content of the image. However, time series analysis may
require atmospheric effects be removed to align images radiometrically, and
simple visual interpretation may benefit from the removal of path radiance (haze).
Below we review the different approaches that can be taken and what is involved
in a practical sense.
Empirical line and dark pixel subtraction—Strictly speaking, Eq. 4.1 is an
approximation as it ignores the more complex scattering pathways and spatial
variation in surface reflectance. Nevertheless Eq. 4.1 is the basis of a number of
atmospheric correction schemes. Importantly Eq. 4.1 establishes a linear relationship at each wavelength between the surface reflectance, L w (k), (what we
want) and the at-sensor reflectance, L t (k), (what we have). The two components
that must be deduced in each band are the scaling factor and offset, t(k) and L a (k).
If reference ground truth reflectances are available then these parameters can be
estimated by linear regression for each band and this is the basis of the so-called
‘empirical line’ atmospheric correction. Smith and Milton (1999) review the
requirements to accurately perform this correction, in particular reference targets
must be homogenous and substantially larger than the pixel size to avoid adjacency or point-spread function (PSF) effects (Milovich et al. 1995). In addition
reference targets should encompass the range of reflectances in each band, to avoid
extrapolation. Empirical line is also suitable for high spatial resolution satellite
sensors (Karpouzli and Malthus 2003). A basic form of empirical line is dark pixel
subtraction (Mather 1999), which assumes the darkest band values in the image
represent a surface reflectance of zero. This is not recommended for most processing algorithms that require atmospheric correction (e.g., inversion methods,
Sect. 4.3.5) and will have little or no effect for most classification approaches
(Caplosini et al. 2003).
Cloud shadow method—Lee et al. (2007) present a convenient method that can
be applied if an image contains deep water areas, clouds, and cloud shadows on
deep water. The method also requires an estimation of the ratio of direct to total
irradiance on the surface, which can be obtained from freely available and simple
to use radiative transfer models such as SBDART (Ricchiazzi et al. 1998) or
libRadtran (Mayer and Kylling 2005). The method is relatively insensitive to this
estimation so while it requires specifying atmospheric constituents a ‘standard
atmosphere’ assumption may be sufficient.
Radiative transfer modeling—A more sophisticated approach is to use a
radiative transfer model, either to evaluate t(k) and L a (k) or to parameterize a more
complicated inversion that more fully captures multiple scattering photon paths.
90
J. D. Hedley