13-1: As will be mentioned
later on this page, prior to obtaining the imagery,
the prime task will be to become familiar with the
major features in the scene with the intent of picking
some specific locations and examples for use as training
sites presuming that classification is your principal
objective. After the classification is done on the
Landsat image, it is essential to return to the field
to check out a number of places where classes have
been defined in order to test the overall classification
accuracy of the computer-based mapping. BACK
13-2: Actually, a strong
case can be made for on-site measurements of any of
the first 8 on this list. Most of these can be measured
automatically by instruments in place. The trick is
to get the resulting information to you, the user.
This is usually simple, but not cheap. The measurement
station can be equipped with a radio transmitter that
relays the data from ground to a satellite and then
to a receiving station (either a central one such
as at NASA or an agency, or to you directly). This
is the basis for the concept of Data Collection Platforms
(DCPs) described later in this section. It is also
behind the idea of Regional (Applications) Centers
described in Section 20. BACK
13-3: Any number of possibilities
come to mind. Yours is probably valid and sensible.
To give an example of what a good answer might be,
try this: The problem is to monitor a region's anticipated
harvest of its staple crop, winter wheat. The platforms
your choose would be a geostationary satellite to
monitor weather conditions, a Landsat to assess growth
stages, and a radar satellite to monitor soil moisture.
Another satellite, with multiband thermal, would also
be helpful. All of these will gather data over time.
As the growing season progresses, observers (county
agents or the farmers themselves) in the field send
crop information. Periodically, you and/or they can
bring portable sensors to get on-the-spot spectral
data. The final result should be a quantitative estimate
of crop status in terms of total yield, quality of
crop, and any disease problems. All of the "multi's"
have been brought to bear on this task. BACK
13-4: Optimal classes for
lower resolution systems are those that are homogeneous
over relatively large areas. Several that come to
mind are large water bodies; clouds; vast forests;
deserts; snow fields. Those that provide poor samples
when resolution is low include residential areas (unless
no attempt is made to define this category in terms
of individual buildings but just as a general class)
and multi-use terrain (e.g., a power plant near a
river with adjacent fields and forests). BACK
13-5: The first thing you
should notice is how variable most scenes are. The
question of signature acquisition has a rather complex
answer. Signature of what? This depends heavily on
your choice of category or class and how general it
can be. If you want a signature of a house, how close
should you be. If too close, you only get part of
any house. If the whole house fills the field of view,
then how meaningful is the signature for a second
house, likely to notably different from the first.
The yards and landscaping will change from place to
place. What about the land behind the house(s) off
in a distance. This varies from one view to the next.
Spectral signatures of the real world are hard to
specify in terms of "typical". Spectral signatures
work well mainly when the target is uniform: such
as a rock outcrop, a field of some particular crop,
a wetlands, a water body. BACK
13-6: The center left pixel
has wheat, corn, and grasslands - all vegetation but
it would be hard to try to extract which class should
be cited. In the center pixel, the lower right pixel
at higher resolution (smaller surface area) is mixed
with five classes: hardwoods, corn, shrubby meadows,
rock outcrop, pines. In this case, even with the improved
resolution, the mixed pixel problem does not go away.
By chance, that pixel straddled a place on the ground
where several classes all met. Analysis of that pixel
in a classification will likely prove a problem, with
an unreliable identification. BACK
13-7: This is a reasonable
pictorial scenario: There will be a cluster of 2,
3, or more actual factories - usually elongate buildings,
with connecting walkways and shipping paths and roads.
There may be a separate power plant. There should
be an administrative or business building, commonly
by itself. There may be landscaping - a lawn, trees,
leading to the plant's frontage. One or more parking
lots surround the complex. There may be a gatehouse,
for admittance. There could be a pond for water supply
or a water tower. There is likely either a rail sideline
or a place where trucks can park and load. Thus, the
category "factory complex" is just that, a congregation
of structures of diverse appearance and use. BACK
13-8: In the overall classification,
the accuracy is simply calculated as the sum of the
number of correct identifications for each class as
a percentage of the total number of units (pixels)
which consists of the sum of the correctly identified
pixels for a class plus the pixels representing errors
of commission and omission associated with that class.
These errors may significantly increase the denominator
to numbers greater than the actual total for any one
class. BACK
13-9: An error of commission
is a measure of the ability to discriminate within
a class and occurs when the classfier incorrectly
commits pixels of the class being sought to other
classes. In this agricultural example, the commission
error for corn stems from improperly calling other
classes corn, so that seven pixels labeled as corn
are really a composite of other classes. An error
of omission measures between class discrimination
and results when one class on the ground is misidentified
as other class(es) by the observing sensor and/or
the classifier. Thus, Landsat fails to recognize and
correctly identify all 43 pixels of corn as such,
and labels 18 of these pixels as other classes. BACK
13-10: A partial list:
Training site mislocated; Inadequate training samples
(number too few and/or sites non-uniform (inhomogeneous);
Mixed pixel effect; Class improperly defined; Ground
truth inaccurate; Signature extension invalid; Stage
of growth not considered; Temporal difference between
ground truth and date of scene overpass by satellite.
BACK
13-11: Viewing conditions
can be controlled; Reflectances can be quantified,
using calibration targets; Different surface can be
examined to get averaged values; Measurements can
be repeated under the same or other conditions; Targets
can be sampled (collected) directly to determine composition
and other properties affecting reflectances. BACK
13-12: Bands at 0.67 (chlorophyll
absorption) and 0.87 µm (cell wall reflectance).
A third channel at 0.56 µm is sensitive to "greenness".
BACK
13-13: As the viewing
angle relative to the surface increases, the reflectance
ratio decreases (in numerator [IR] greater than denominator
[red]); variations to a lesser degree occur as the
direction of look at the sample varies relative to
north. BACK
13-14: The normalization
(use of calibration targets) applied to the field
spectra removes certain external influences, such
as atmospheric absorption, and provides a truer indication
of the natural reflectances of the materials themselves.
(In lab conditions, the samples are illuminated with
air around them [a vacuum could be used, but usually
is not], so that water and gas absorption troughs
are not eliminated.) BACK
13-15: 1) A standard reflectance
target could be set up, with a small radiometer aimed
at it, so that when the satellite is passing over
the scene, a ground-based value for reflectance under
the general illumination conditions acting on the
control area within the scene can be measured; 2)
seismometers, for detectung earthquake precursors;
3) a floating platform that gives sea surface temperatures;
the platform can be anchored or it can float, with
its position determined by GPS. Many other possibilities
could be mentioned. BACK
13-16: Hyperspectral remote
sensing provides a continuous, essentially complete
record of spectral responses of materials over the
wavelengths considered. For relatively "pure" materials,
e.g., individual minerals or tree species, it is possible
to construct a spectral curve from the hyperspectral
data that can then be matched with spectral signatures
of individual materials collected from laboratory
or field measurements and available in data banks.
Specific reflectance peaks and absorption troughs
can be read directly from these curves to allow precise
identification of a material, class, or feature. With
careful analysis (using Fourier procedures), mixtures
of two or even three different materials, etc. can
be identified as the components of the compound spectral
curve. BACK
13-17: First, a distinction
should be made between "skylight" and "sunlight".
The latter refers to all the radiation entering the
Earth's atmosphere from the Sun and includes wavelengths
extending from the high frequency UV to the lower
frequencies in the infrared. Skylight is a term that
signifies the scattering of sunlight by atmospheric
gases and particles; scattering occurs to varying
degrees at different wavelengths. Rayleigh scattering,
causes by the gaseous molecules in the atmosphere,
affects preferentially the shorter wavelengths. These
include the blue, so that the sky appears to be the
source of that color since, in fact, it is - the molecules
selectively scatter blue but pass longer wavelengths.
Sunsets are red because that is the color remaining
after the shorter wavelengths (through greens) have
been further scattered (by Mie scattering - from dust
particles, etc. - in addition to Rayleigh scattering),
plus the effect of a lower Sun angle (near the horizon)
that increases the atmospheric path length. The sky
from the Moon is black - that is just the result of
there being no atmosphere at all to scatter the sunlight.
Thus, sunlight illuminates the lunar surface but not
anything in the near-vacuum of space beyond. BACK
13-18: Three absorption
bands, at 1.3 - 1.5 µm, 1.8 - 2.0 µm,
and 2.5 - 3.0 µm, should be avoided whenever
remote sensing is conducted through the atmosphere
- unless the objective is to study atmospheric properties
themselves. BACK
13-19: A paired absorption
band for Goethite near 6 µm is distinct from
the single band for Hematite at 7 µm. BACK
13-20: When the four plant
types are analyzed in a spectrometer, only a small
part of the plant is taken as a sample. It is likely
that each of the types will have its own characteristic
leaf or frond shape and that, taken as a whole, the
overall appearance of any type will differ geometrically
from most other types. Thus, oat hay and potato as
crops are clearly dissimilar in the way they look
in bulk. Thus, the combination of spectral response
and diversity of shape will produce slightly different
signatures, mainly in the depth of any absorption
features. BACK
13-21: Probably not -
the band is too broad. Distinction and identification
is much better in the 8 - 12 µm thermal interval.
Of course, a hyperspectral sensor can separate the
two rock types, based on their principal mineralogy,
in the 2.3 µm region. BACK
13-22: Yes, the visual
differences also show up as absorption bands in the
spectral curves. They would be distinguishable. BACK
13-23: The Bronzite's
absorption bands will affect TM Bands 4 and 5; the
Diopside doesn't have bands in these spectral intervals
but possesses a moderate absorption band that will
reduce Band 7 reflectance. BACK
13-24: As the Aluminum
in octahedral coordination increases, the absorption
band near 2.2 µm shifts towards slightly longer
wavelengths. BACK
13-25: Two factors control
this effect: the larger grains allow more absorption
and the smaller grains provide a higher proportion
of surface area available as reflectors. BACK
13-26: For n = 1, sin
theta = lambda/d. d = 1/5000 lines/cm = 2 x 10-4
cm/line = 2 x 10-6 meters. For red
light, sin theta = 650 x 10-9 meters/2
x 10-6 meters = 0.325; theta therefore
is 19°. For blue light, sin theta = 450 x 10-9/2
x 10-6 = 0.225; theta is 13°. The dispersion
in the visible is therefore 6°. The wavelength
continuum in this spectral range (and beyond) is therefore
spread out, and separated, to be either recorded on
a photographic plate or by light-sensitive detector/counters.
BACK
13-27: The detector interval
is 10 nanometers, or 0.01 µm, which is a small,
but still finite interval in the spectral continuum.
10 nanometers = 100 angstroms; a discrete emission
line on a photographic plate is quantified as a fraction
of a single angstrom, e.g., a diagnostic line for
sodium (Na) when excited in an emission spectroscope
is at 5889.953 angstroms (in the yellow-orange, the
color of a sodium-vapor lamp used in street lighting).
BACK
13-28: Each of the Landsat
bands, either MSS or TM, are spectrally broader. Thus,
the green MSS Band 5 actually includes some colors
that contain bluish or yellowish contributions. They
are not as pure, so that color composites may not
be as precise. Thus if a red object on the ground
is rendered in a TM natural color image, its particular
shade of red may contain some yellow tones. In AVIRIS,
if the individual channel chosen for the red component
in the composite is close to the "true" red color
of some object, its representation in the AVIRIS image
will nearly match that of its actual shade. However,
overall, the AVIRIS and Landsat natural and false
color composites appear similar because the range
of colors in nature is usually greater than those
discriminated at the 10 nanometer level. BACK
13-29: Very little direct
evidence. One might miss the abundant knowledge about
this diverse alteration if only Landsat type images
were examined. However, Landsat TM bands 5 and 7 will
likely show at least some of this alteration. But,
these AVIRIS images of Cuprite clearly demonstrate
the power of being able to select channels that lie
close to the peaks and troughs of spectral curves
rather than lose some essential detailed information
when these curve features are subsumed into broader
bands. BACK
13-30: Nothing mapped
means the computer classification could not match
the material that is rendered black with any of the
pre-selected signatures. It may be another crop type
not designated by a name (i.e., was not picked during
ground truthing) or is a soil variant. BACK
13-31: The hot vents (in
blue-white) show up better in the thermal image. The
areas in red are probably basalt or andesite flows
that were laid down during one or more recent eruptions
(Etna is the most active of the Mediterranean volcanos.)
BACK
Collaborators: Code
935 NASA GSFC, GST,
USAF Academy,
Webmaster: Bill Dickinson Jr.
Primary Author: Nicholas M. Short, Sr.
email: nmshort@epix.net
Contributor Information
Last Updated: July '99
Site Curator: Nannette Fekete
Please direct any comments to the Site
Curator eerstweb@gst.com.