VI. Satellite to Satellite
(S2S) Interpretation of
Once validation boxes have been
located and entered into the geodatabase,
interpretation of changes can begin. S2S
interpretation is based on the principles
applied to airphoto interpretation. The
human brain brings in spatial context to
help interpret patterns of shape, color,
and text to determine what is happening
in an image.
The differences between aerial
photo and satellite interpretation are the
scale of the analysis and the spectral
depth of the respective image types. In
high-resolution aerial photos, individual
elements on a landscape (i.e. trees,
small patches of rock, or small streams)
are interpretable at a scale familiar to a
human. In satellite imagery at the scale
of Landsat TM/ETM+, such detailed
features individually are not discernible.
This requires some degree of retraining
for interpretation. Satellite imagery
provides an important advantage over
aerial photos in its increased spectral
depth. Landsat imagery provides spectral
information about the surface in the
near infrared and midinfrared regions,
both of which are extremely useful for
distinguishing different cover types. The
tasseled-cap transformation captures the
variation in the six visible, near-infrared,
and midinfrared bands of TM with just
three axes, allowing display on standard
The S2S process also requires that the interpreter
understand how to interpret spectral patterns in images of the
mathematical differences between two tasseled-cap images.
Table 3 provides an interpretation key for difference imagery.
This assumes that the image is displayed in a viewer with
brightness in red, greenness in green, and wetness in blue. See
SOP 2 if this has not been set as the default for viewing.
The interpretations in table 3 provide a first
approximation of the likely change that has occurred. The
interpreter must then apply knowledge of the conditions
before and after the change (using direct interpretation of the
tasseled-cap images), knowledge of the ecosystem, and spatial
context to further hone the interpretation. Finally, reference to
a single-date, high-resolution digital orthoquad can improve
understanding of the potential land-use condition in which
Table . Interpreting tasseled-cap difference images when early image is subtracted
from late image.
Direction of tassled-
early to later
Dramatic loss in vegetation, either because of
disturbance or because of phenology; OR presence
of cloud in later date.
Subtle increase in
Often associated with loss of vegetation from areas
that were very bright and green at early date—a
good example is a disturbance in shrub field or
young hardwood/conifer plantation
Increase in all
Increase in vegetation, especially conifer. Common
in clearcuts that were vegetated partially in early
date and which are increasing in HW and conifer;
OR an increase in snow from a relatively vegetated
starting point; OR presence of cloud in early date
and conifer in later date.
Increase in wetness,
Increase in conifer percent cover. Common in
clearcuts transitioning from hardwood or shrub
dominance to conifer dominance; OR increase
in shading in later date (take into account date of
image); OR presence of cloud in later date; OR
increase in water (channel changes).
Typically, increase in broadleaf component, either
shrub, grass or hardwood tree, caused either by
recovery from disturbance or by phenology; OR
presence of cloud shadow in the early date and
vegetation in the later date.
Bright green Increase in
Increase shrub/grass/hardwood cover from a
condition that was bright before, but not green—
typically found when river beds come back to
vegetation, but also common where snow was
present before and now is gone.
digital monitors in the three color guns (red, green, and blue).
Color-infrared photography, when available, could provide
insight into the near-infrared reflectance of the surface, but
not typically into the midinfrared reflectance. The added
spectral depth and interpretability afforded by the tasseled-
cap transformation of TM/ETM+ imagery requires further
training to interpret, but increases the ability of an interpreter
to describe conditions.
Because of the need for training, the authors have
provided an image library to be used in conjunction with the
S2S interpretation process. Before S2S interpretation begins,
interpreters should use this library to gain familiarity with
the main cover types and their representations in tasseled-cap
space. The library associated with this version of the protocol
SOP . Validation of Change-Detection Products 101
Table . Satellite-to-satellite disburbance agents.
Fire or insect.
Table . Certainty scoring for change classes.
[Score is assigned by the relative degree of agreement with the conditions
statement, with zero indicating that the condition is not found. Change call:
Total certainty score ranges from 0 to 4; Disturbance agent: Total certainty
score ranges from 0 to 3]
Spectral change vector is distinct from change
vector of similar starting types in surrounding
0, 1, or 2
Area of spectral change is large and consistent
0 or 1
Spectral condition of endpoints is interpretable
and consistent with change call
0 or 1
Shape is consistent with disturbance agent
0 or 1
Size is consistent with disturbance agent
0 or 1
Landscape position and context are consistent
with disturbance agent
0 or 1
interpretation is occurring. A final change call is made and
categorized into one of 15 categories, listed in table 2. This
table can be printed out separately to provide an easy reference
when onscreen editing is done.
In addition to determining the change class, the
interpreter makes a call about the disturbance agent (table 4).
This is drawn from the combination of the spatial context and
the observed spectral change.
In assignment of both change class and disturbance agent,
the interpreter may have varying degrees of certainty. Degrees
of certainty can be quantified using simple rules that describe
the elements of the decision-making process when a change
call and disturbance agent are identified. Table 5 lists the
conditions and scoring for certainty scores in the change calls
and the disturbance agents.
A variety of factors could lead to false positives in the
automated change algorithms, and these must be filtered
out by the interpreter. To achieve this, the interpreter must
be aware of the factors that can lead to false positives in the
automated algorithms. The following is a partial list of four
major effects that the interpreter can take into consideration
when evaluating the spatial patterns of spectral change.
Even with excellent image-wide geometric registration
of two images, local distortions can remain. These are
mostly inconsequential, but if they occur at a boundary
with sharply contrasting spectral types, even a small
geometric shift can result in large apparent shifts in
spectral value. Thus, misregistration will manifest
itself most strongly around small patches and linear
features that contrast with their surrounding matrix
(i.e. avalanche chutes, rivers, or small openings
in the forest). A difference image will show that
large change has occurred, but examination of the
two source images separately will show that the
shape and color of the area has not changed. This
indicates that misregistration has caused the spectral
change. Another clue comes from the shape of the
apparent spectral change in the difference image. If
the difference is only at the margins of a small patch,
and the complementary spectral change occurs on
the other side of the patch, then misregistration also
is implicated. In cases where it is unclear whether
misregistration plays a role, indicate the change type
but give a correspondingly low confidence score.
After the summer solstice, the sun is lower at the time
of each successive Landsat image acquisition, so the
shadows get longer. Landsat images are acquired at
approximately10:30 a.m. local sun time, meaning that
the sun is approximately in the southeast, although
the azimuth of the sun also varies with the seasons.
The illumination variability caused by the different
angles of the sun is sometimes subtle, but can cause
some shadowing effects on the northwest aspects of
slopes. For example, a small opening in the forest in
1 year may appear to go away in another year, simply
because the patch has fallen into shadow in the latter
date. This can happen even with a difference of just
a week or 10 days in the date of image acquisition,
especially as days change more rapidly in late August
and September. If it seems likely that shadow may have
caused difference, do not include it.
102 Protocol for Landsat-Based Monitoring of Landscaped Dynamics at North Coast and Cascades Network Parks
The decision is not always straightforward.
Illumination will cause bulk changes in reflectance for
entire hillsides, but so can other effects. For example:
Insect infestation can affect entire hillsides in forested
If it is difficult to tell, label the change, but score
the certainty appropriately. In this case, because the
spectral condition is not interpretable easily and
because the signal may or may not be similar to other
conditions, table 5 would indicate a certainty score of
Year-to-Year Phenology Changes
Because the date at which cloud-free imagery can
be acquired varies from year to year, and because of
climatic variations between years, the phenological
state of deciduous vegetation often varies from image
to image. This is true especially at the alpine/subalpine
interface, where snowpack duration drives much of
the timing from year to year. Changes in phenological
state will have spectral manifestations that, although
corresponding to an actual change in the state of the
vegetation, do not represent a change in cover type
or quality that is of interest in this work. Automated
spectral change-detection methods will be fooled by
this spectral change, but a human interpreter can often
account for such phenological change by comparing
the spectral changes in deciduous vegetation in one
location with the overall spectral changes of similar
vegetation in other similar parts of the landscape.
If spectral changes are definitely associated with
phenology, change should not be labeled. If the
change call is difficult, the certainty score should
be adjusted accordingly. If all of the vegetation in a
given type is experiencing similar changes, then the
first change-call condition in table 5 is not met, so the
maximum certainty score is 2, with likely lower scores
because the other conditions are not met.
Problems With Radiometric Normalization
Radiometric normalization with the MADCAL
algorithm appears to be quite robust, but no automated
normalization process is perfect. The magnitude of
errors in the fit between the two image years often
scales with the brightness of the target. Thus, small
errors in radiometric normalization can lead to spectral
differences that are quite evident in the difference
imagery used for interpretation. When the two end-
point images also are referenced, however, there is
often no apparent difference in the condition, because
each image is displayed according to its own display
scaling equation that compensates for the scaled error.
Thus, it is often possible to rule out the apparent
changes that are caused by radiometric error.
VII. Imagery Setup and Satellite-to-
Satellite Interpretation Process
The actual setting up of imagery and S2S interpretation
process is as follows:
1. Create a tasseled-cap difference image for interpretation.
a. In ERDAS Imagine, do the following steps:
i. From the main icon panel, click the Image
(1) This is the icon with the magnifying glass
over a small raster box.
ii. Select Utilities/Operators
(1) The Two Input Operators dialog window
will pop up (fig. 8)
(2) The “Input File #1:” is the LATER date
image (the changed image).
(a) Use the entire study area, tasseled-cap
image used in SOP 3 for the change
detection (not the aspect-subsetted
Figure . Window used to create difference images.
SOP . Validation of Change-Detection Products 10
(3) “Input File #2” is the EARLY date image
(the baseline image).
(4) The output file should be named to indicate
both parent images, and should include the
word “minus” in it, to clearly indicate that it
is a difference image.
(a) This image is the “Difference Image”
referred to subsequently in this section.
(5) Change the “Operator:” to the minus sign.
(6) Change the output type to Signed 16-bit.
(7) Click OK.
2. All of the interpretation is done in ArcMap.
a. Ensure that the following layers are present in the
i. The baseline tasseled-cap image (the baseline
image was defined in SOP 3).
ii. The changed tasseled-cap image (the image from
the later year).
iii. The difference image.
iv. A DOQ specific for the area being interpreted.
(1) Install the Terraserver add-in to ArcMap.
(a) Terraserver provides free access to
digital orthoquads that will be used to
aid the interpreter to understand landuse
in an area.
(i) If Terraserver is not already installed
in ArcMap, see information at the
ESRI web site for instructions.
(ii) Generally, the steps for ArcGIS 9.0
are as follows:
1. At the ESRI support website
(support.esri.com), click on
downloads and search for
2. Download and install according to
3. This will involve the .NET
architecture, which, if not already
installed, also will
need to be installed.
4. Once Terraserver is installed, it
can be viewed within ArcMap
v. The feature class that holds the validation boxes.
(1) Select this layer and then enable editing on
this layer, if not done already.
(a) Select Editor/Enable editing.
b. Begin with the first validation box where
interpretation is to begin (created either by method 1
or 2). It likely is that a systematic interpretation from
top to bottom or left to right in the study area will be
i. Open the attribute table in the feature class that
holds the validation boxes.
(1) Locate the current validation box in the
(a) The selection tool in ArcMap (a small
white arrow next to a square with three
small polygons inside) can be used to
click and drag over the area of the box,
(2) Fill-in the information on the box itself.
(a) Label the box with a unique numerical
sequential ID, beginning with 1.
(b) There is no event ID for the box, which
(c) Assign a change type of 1 (no-change;
see table 2) by typing directly into the
“change_type” cell for the row of this
box, if the number is not already there.
(d) Assign a certainty score of 4.
(i) This sets the background to the no-
c. With the difference image as the visible layer
(checkmark in the Layers window), methodically
work through the entire 1.5 km by 1.5 km box
looking for spectral evidence of change.
i. In the difference image, no-change is shown with
ii. Bright and colored areas in the difference image
indicate large spectral change between years.
Examine each patch of such potential change.
(1) Zoom into the area with the potential
(2) Turn off the difference image by unchecking
the box. Note which of the two tasseled-cap
images is now visible – it will be the image
that is higher up in the list of the layers.
(3) Then turnoff that image by checking its box
10 Protocol for Landsat-Based Monitoring of Landscaped Dynamics at North Coast and Cascades Network Parks
Documents you may be interested
Documents you may be interested