Computing Radiances, Reflectance and Albedo From DN'S: (September 21, 2005)
Computing Radiances, Reflectance and Albedo From DN'S: (September 21, 2005)
1. Introduction:
In some cases, but not all, it is preferable to convert raw short wave satellite image data
to physical quantities before using the data to interpret the landscape. Important physical
quantities include radiance (at-ground or at-satellite), reflectance, or albedo. For
example, it is the at-ground reflectance that is characteristic of a particular surface type,
and independent of the illumination and atmospheric characteristics. In the sections
below we discuss several approaches to the development of physical data layers.
With Landsat for example (see the online Landsat Science Data Users Handbook) the at-
satellite radiance must be computed from the instrument calibration function
I λ = aDN + b (1)
where a and b are the gain and offset. Each spectral channel will have different gain and
offset values. Usually the radiance given by such a formula will have units Watts per
square meter per steradian per micron (i.e. Wm −2 sr −1 µm −1 ).
3. Solar Illumination
In many situations, the incoming flux is dominated by a bright narrow beam of radiation
coming directly from the sun. The strength of the solar beam at each wavelength is the
“Solar Spectral Irradiance” ( S λ ). For example, the value for band 1 wavelength for ETM
is about S λ = 1969 Watts per square meter per micron. Note that this “irradiance” is not
angularly resolved (i.e. it is not “per radian”). The angular width of the beam is about
0.5244 degrees (about 31 minutes) of angle, but this width is not usually needed for the
computations below. The solar irradiance value outside the earth’s atmosphere changes a
bit with season as the earth’s orbit is slightly elliptical. The irradiance is inversely
proportional to the square of the earth-sun distance. If the earth sun distance is increased
by 1%, the irradiance decreases by 2%.
The illumination of the earth’s horizontal surface will depend on the solar zenith angle
( φ ) according to
Fλ = S λ cos φ . (2)
As the sun sets lower in the sky, the illumination decreases, reaching zero at sunset. If
the local surface is tilted, the illumination should be computed using not the “zenith
angle” but rather the angle between the “surface normal vector” and the sun. Note that
with a complex forest target, the illumination of particular leaves may be as great as S λ or
4. Lambertian Surfaces
When this solar beam hits a complex surface, the reflected radiation must in general be
described by a radiance field, that is, an angularly resolved field of radiation. It is not a
narrow beam anymore. The general description of how a complex surface reflects
radiation is the Bidirectional Reflectance Distribution Function (BRDF); the ratio of the
reflected radiance to the incident irradiance.
R(θ , φ , θ F , φ F ) = I λ (θ , φ ) / Fλ
where dΩ are the increments of sold angle covering the hemisphere. If the radiance is
independent of zenith angle, the integral becomes
where rho is the spectral reflectance. Solving (5) for reflectance and using (2) gives
ρ λ = πI λ / Fλ = πI λ / S λ cos φ (6)
This is a familiar formula seen in textbooks and in the Landsat Science Data Users
Handbook. Sometimes the earth-sun distance correction is included. If we know the solar
spectral irradiance and we measure the reflected radiance, (6) gives us the spectral
reflectance.
To judge for yourself the validity of the Lambertian assumption, place a piece of white
paper on a table. Look at the paper from different angles. Does its apparent brightness
change?
Consider for a moment, our reasoning with point c and d. If an object is bright (e.g. a
snow field), the reflected radiation will be intense and scattering out of the path will
exceed the scattering in. Conversely, if an object is dark (e.g. a forest), the reflected
radiation will be weak and scattering in will exceed scattering out.
As it turns out, in the visible range where most scattering occurs, most earth objects are
dark (e.g. soil, water, vegetation). The exceptions are snow (80%) and bright sand (50%).
Furthermore, the haze that causes strong path radiance is highly variable from day to day.
Thus, if not corrected for, it will degrade any change detection study.
Path radiance is familiar to us (see Figures 1,2,3). It prevents us from seeing distant
objects on a hazy day. It gives distant mountains a pale blue appearance.
at each wavelength ( subscript lambda deleted here). The path radiance will depend on
the strength of the illumination and the density of scattering particles in the field of view.
It will be a decreasing function of wavelength, because shorter waves are scattered more
than long waves. Because of this trend, it can often be ignored in the near infrared,
unless the haze is very thick with some particles as large as 1 micron.
As we will see below, path radiance can be estimated and removed using dark object
subtraction (DOS), if the haze layer is uniform over the scene. This assumption is
Another source of path radiance is a layer of thin cirrus ice cloud in the upper
troposphere. These layers are seldom very uniform however, so path radiance removal
could be problematic.
6. The effect of illumination, gains, offsets and path radiance in remote sensing
It is sometimes supposed that sensor calibration aspects or path radiance will not cause a
problem if we always use band ratios, contrast enhancement or classification methods.
Let’s investigate these claims (Song et al., 2001).
a. NDVI
Consider a first a normalized quantity such as NDVI, expressed in terms of at-surface
reflectance:
ρ 2 − ρ1
NDVI = (8)
ρ 2 + ρ1
Such a quantity could be determined using a portable spectrometer. Can this value be
determined using satellite data? What quantities should be entered in the formula: DNs,
at ground radiance, at-satellite radiance? If we use (1) and (6), canceling a pi everywhere
ρ 2 − ρ1 (a 2 DN 2 + b2 + I PR 2 ) / F2 − (a1 DN 1 + b1 + I PR1 ) / F1
NDVI = = (9)
ρ 2 + ρ1 (a 2 DN 2 + b2 + I PR 2 ) / F2 + (a1 DN 1 + b1 + I PR1 ) / F1
Using (9) we can identify the special circumstances in which DN values can be used to
obtain quantitative NDVI values: no path radiance or offset and equal gains and
illumination. Only in this unlikely case does NDVI from (8, 9) equal
DN 2 − DN 1
NDVI = (10)
DN 2 + DN 1
A less restrictive condition is that the NDVI of an identical surface should stay constant
from date to date. Such a definition would at least allow NDVI change detection. If
NDVI is defined in terms of calibrated at-satellite reflectivity (Isat) using (7)
Now if we assume that Isurf and IPR are proportional to the illuminating irradiance, that is
I surf 2 = ρ 2 F2 and I PR 2 = h2 F2 etc., we obtain
where “h” is some measure of the haze density and scattering efficiency. According to
(12) the definition of NDVI (11) gives invariant values under changing illumination but if
“h” were to change, the NDVI value would change also. If the path radiance can be
neglected (e.g. h = 0, a clear day) (12) returns to (8).
A special problem arises with clouds. In a cloud or mountain shadow for example, the
NDVI may be shifted (reduced) because the skylight illuminating the pixel is richer in
blue light, and very deficient in NIR.
b. Contrast stretch
Can a procedure of contrast stretch correct for unknown illumination, gain, offset and
path radiance? Yes, in principle! Any linear transformation on radiance (such as adding
path radiance (7)) or in converting DN to Radiance value (gain and offset (1)) can be
easily undone by the appropriate linear contrast stretch. Thus, for example, an image
taken in a hazy atmosphere can be restored by contrast stretching to appear like an image
without haze. In making this claim, we assume that the radiometric resolution and
saturation value of the sensor is sufficient to prevent permanent loss of information by the
original transformation. Also, we usually adjust the contrast stretch by eye, instead of
using the specific values that would compensate for environmental or instrument factors.
c. Classification
Now we consider the impact of a linear transformation on classification (Song et al.,
2001). As a linear transformation will only shift and stretch the points on a scatter
diagram, the results of many classification algorithms will not be thereby altered. As an
example, consider a “maximum likelihood” algorithm. As it normalizes the Euclidean
distance between pixels with the variances computed in the same Euclidean way, the
result of the classification will not be altered by a linear transformation (i.e. shift or
stretch).
When at-ground radiance or reflectance data are not available, references areas can
sometimes be used to develop quantitative remote sensing results. For example:
for the dark pixel. Defining I dark = I sat for that pixel, the other radiance values in the
scene can be corrected using
DN − DN D
ρλ = (15)
DN W − DN D
To derive (15), use (1),(6) and (14) to write the reflectance for all three pixels
πI π π π
ρ= = ( aDN + b − I PR ) ρ D = (aDN D + b − I PR ) = 0 ρ W = (aDN W + b − I PR ) = 1
F F F F
The second equation gives b − I PR = −aDN D . Subtracting the second and third equations
gives a( DNW = DN D ) = F / π . Applying these two expressions to the first equation gives
(15). Note that if the two reference pixels were not perfectly black and white, but their
reflectances were known, a similar formula to (15) could be found.
In practice, it is difficult to find a pixel with a perfect reflectance. The best choice in the
visible range would be a deep dense cloud. Clean snow might work too. Its reflectance
can exceed 90% in the visible, but it drops off quickly in the NIR. It’s difficult to find any
natural surface that has a reflectance exceeding 70% in the NIR, so (15) will probably
work better in VIS than in NIR. A problem with Landsat (and sometimes MODIS) is that
the sensor will saturate (e.g. DN=255) for highly reflective pixels. These pixels are too
bright for the sensor. Thus no pixel with near perfect reflectance exists in the image.
If we define the two invariant DN values as DN1 and DN2, then a scaled DN value is
If DN2>DN1, a pixel brighter than pixel 2, will have a scaled DN value greater one. A
pixel darker than pixel 1, will have a scaled DN less than zero. When the layer of scaled
DN values is computed for each scene, they can be compared to detect change. Because
of our method, the scaled DN values at the two reference pixels will be unchanged
between the two dates (0 and 1 respectively).
Additional methods for change detection are given by Song et al., 2001.
8. Albedo
The albedo is an important quantity in climate theory. It is defined as the ratio of the total
reflected irradiance to the incident irradiance
For climate applications, we would like to determine the global patterns of albedo using
satellite data. This is difficult for two reasons (Liang and Strahler, 1999). First, satellites
only measure reflectance in a few narrow spectral bands. For albedo, we need the
reflectance at all wavelengths where the sun shines. Second, satellites usually measure
reflectance at one angle incidence only, and at one reflected angle only. If the surface
were Lambertian, this one angle would be sufficient. The assumption of a Lambertian
surface is not sufficiently accurate for some albedo computations however.
A third problem can arise in relation to angle. In the real world, the angular nature of
illumination can vary from hour to hour. If the sky is clear, the incident radiation hitting
the earth’s surface will be the sun’s direct narrow beam. If the sky is cloudy, the incident
radiation will be diffuse, almost isotropic. In this latter case, satellite can never be used
to observe the reflective process.
All is not lost however. The Bi-directional Reflection Distribution Function (BRDF) that
describes the reflectance as a function of incident and reflection angle, is a constant
property of the local land surface. It is invariant under changes in lighting conditions. If it
can be deduced from multi-angle satellite measurements during clear sky days, the total
albedo can be computed for any angular distribution of incident radiation. For archiving
purposes, it has been found useful to define to special albedos.
If these two albedo fields are known globally, the actual albedo for any meteorological
situation can be determined by interpolating between them (Schaaf et al., 2002).
9. References
Liang, S., A.H. Strahler, C. Waltall, 1999, Retrieval of Land Surface Albedo from
Satellite Observations: A simulation study, J. Appl. Met., 38, 713-725
Schaaf. C.B. et al., 2002, First operational BRDF, albedo nadir reflectance products from
MODIS, Remote, Sens. Envir., 83, 135-148
Song, C., et al., 2001, Classification and change detection using Landsat TM data: When
and How to correct atmospheric effects, Remote Sensing of Environment, 75, 230-244
10. Figures
Figure 1: A typical summer afternoon in northern Maine (August 2005). Note the two
mountains in the distance. Why are the mountains lighter and bluer than the trees on the
shoreline?
Figure 3: A small cloud and its shadow in Arizona, as seen from an aircraft. Give two or
three radiative processes that produce some radiance (to the observer) in the shadow.