IMAGE INTERPRETATION

With Guidelines and Suggested Workflow for High-Resolution, Ortho-imagery

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

INTRODUCTION
Image interpretation is the use of certain image attributes, together with knowledge of the objects and
conditions imaged, to identify features and judge their significance. Our brain-eye system is unsurpassed
at interpreting high-resolution imagery—if we first possess the ground-based knowledge about the
resources we wish to measure, or map. With a small investment in
your time you can become a skilled image interpreter—but, you
must first have that ground-based resource knowledge.

A skilled image interpreter will discriminate features of interest

and their attributes from other features in the imagery. For
example, an image interpreter would use their experience to
discern between species and size classes of trees of interest. To be
successful he or she would need to relate their ground knowledge
to the imagery. The interpreter must spend significant time on the
ground with the imagery in hand!

While this document provides useful tips and techniques for

interpreting imagery, the only way to become a skilled image
interpreter is to take your imagery into the field and learn to
bridge the gap between what you know and see on the ground to what you can see on the images. A
suggested workflow to build image interpretation skills—and interpret imagery—is provided in Appendix 1.

IMAGE CHARACTERISTICS—THE FUNDAMENTALS OF IMAGE INTERPRETATION

An image interpreter will use elements that are inherent in imagery to recognize and identify objects of
interest. These image characteristics are comprised of seven elements that we use to derive information
about objects in the image. These image characteristics are: size, shape, tone/color, texture, shadow,
association and pattern1. While it is beneficial to understand these image characteristics, most image
interpreters will use them subconsciously in practice. A description of each image attribute with examples
is provided on the following pages.

The relative importance of each of the seven image characteristics is not constant. The importance of any
characteristic can depend significantly on the scale of the imagery and the properties of the feature of
interest. For example, the shape of a tree crown can be a very important indicator of tree species on high
resolution imagery but, on coarser resolution imagery, individual crowns may not be easily distinguishable
and texture may become a much more important image characteristic.

1
Some experts also identify other image attributes such as Place/Location and Height. While these are legitimate image char-
acteristics, we’re not considering them in this document.

1
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Size
Size is usually evaluated by looking at objects that the interpreter may be familiar with and comparing
their relative size to less familiar objects. For example, an interpreter might look at the discrete
vegetation features in the red rectangle (below) and interpret those features as trees. However, by
comparing the relative size of the vehicles and the one-lane road, in the yellow rectangle, with the
vegetation features, it becomes obvious that the trees are actually shrubs (in this case, sagebrush). If
the scale of the imagery is known, objects can be measured directly—rather than relying solely on
relative size.

Size. Size is usually evaluated by looking at objects that the interpreter may be familiar with and comparing
their relative size to less familiar objects.

2
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Shape
Every imaged feature has a shape. However, shape is not
necessarily diagnostic for interpretation (i.e., the same type
of feature may have different shapes, and different features
may have the same shape). Man-made features, such as the
vehicles, trailers, and roads (previous page), tend to be more
rectilinear than natural features. Natural features, such as
the trees and shrubs in the image have more subtle shapes.
Referring to the image at right, the shape of the tree crowns
is indicative of the type (deciduous/conifer) or even species
of tree. Note: some features like tree crown shape can be
important characteristics on high-resolution imagery but
Shape. Every object has a shape but the
become less important on coarser imagery (and texture
shape may not be diagnostic. However, if
becomes more important).
individual tree crowns are visible in the im-
agery, they can be important clues to type
or species.
Tone/Color
Tone (in the panchromatic image) allows easy distinctions between roads, forests, harvest areas, water
and other features. Color (lower-right image) allows easy distinction between coniferous trees,
deciduous trees (in yellow fall colors), senesced grasses and road surface types.

Tone/Color. Different features have different reflectance properties—light toned areas reflect more
electromagnetic energy (sunlight) than dark features—and thus appear differently on the imagery al-
lowing the interpreter to tell things apart on the imagery. The human eye can discern far more colors
than shades of grey thus; color imagery is usually preferred for interpretation.

3
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Texture
Image texture refers to the apparent roughness and smoothness of an image region created by the
frequency of tonal change on the image. Texture can offer the interpreter clues about the density, age,
and type of vegetation present. Note that texture interpretation becomes increasingly important as the
resolution of imagery becomes slightly more coarse.

Coarse, medium, and fine textured areas.

Texture. Coarse, medium, and fine texture areas. Coarse area represents larger and older trees with canopy
gaps. Fine texture usually connotes young, even-age trees.

4
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Shadow
Imagery is usually acquired within a few hours of solar noon in order to minimize shadows. Shadows
obscure features and are a detriment to interpretation of features within shadow. On the other hand,
shadows cast by objects in the image can give the interpreter information about the shape and size of
those features. It can also provide an indication of the relative height of features. For shadows to be
most effective, the entire shadow should be visible in the image (as in the example below). In some
cases image acquisition might be timed to ensure shadows are present to enhance interpretability of
features that may be difficult to see otherwise (e.g., a fence line in open rangelands).

Shadow. Shadow is an especially important clue in the interpretation of objects. Shadows of buildings, trees,
and other objects reveal characteristics that are not obvious from the overhead view alone. Edges, such as for-
est boundaries, often have characteristic shadows.

5
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Association
By observing objects in the image and observing the objects that are around objects, the interpreter can
make inferences about what the objects really are. In the example below, the presence of a dam (red
ellipse) suggests that this water body is a reservoir. Without that association, the water body cannot be
further classified as a reservoir.

Association. Association refers to the typical co-occurrence of certain features. Vegetation species are often
associated with one another or with topographic characteristics (slope, elevation, aspect, etc.) Complex fea-
tures such as campgrounds, forest management activities, and more, frequently have associated components
that, individually may lead to one conclusion, but considered together lead to a more informed interpretation.

6
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Pattern
In remote sensing, pattern is a recognizable repetition of particular shapes. In the example below, the
checkerboard pattern of rectangular shapes suggests that this once forested area has been harvested
along regular boundaries. In the example on the previous page you can see that the drainage pattern
appears rectangular while in the example below the drainage pattern appears dendritic.

Pattern. Pattern refers to the arrangement of individual objects into distinctive, recurring forms that permit
recognition. Note that different patterns may become more or less apparent at different viewing magnifications.

7
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Change over Time

In addition to these seven image characteristics, when two sequential sets of imagery are available, we
can also add these additional image characteristics: 1) discernible differences in any of the above
characteristics between images acquired at different times; 2) presence or absence of features in
successive imagery; and 3) change of location, position or extent of features.

Convergence of Evidence
Convergence of evidence is the mental (but usually subconscious) summation of all of the interpretable
image characteristics that are present in the imagery to derive a confident interpretation of an imaged
feature. In other words it is the bringing together of several types of information so that a conclusion
may be drawn in light of all available data. It is essentially what an interpreter infers from all the image
characteristics, together with field site data and value judgments derived from previous training and
experience. Interpretation is then made on the basis of what the interpreter feels are the overridingly
important criteria. It is this summation of all of the interpretable image characteristics that makes this
an inherently inductive logic (subjective) exercise that is not easily replicated in a deductive logic
(objective) algorithm in an image processing program. This difficulty has made automation of the image
interpretation process elusive for pattern recognition software.

If the interpreter considers all of the image attributes that are present for a given feature, and all the
attributes of that feature lead to the same interpretation, the interpreter will have confidence in the
interpretation (the evidence converges on the same conclusion). Misinterpretation increases when the
interpreter does not consider all available image attributes or, when the interpretation(s) from the
individual image attributes conflict (an example of a potential misinterpretation is described in the size
characteristic section in the previous section).

Interpretation is an Art
Remember, there is a difference between interpretation and identification. Identification assumes
objective accuracy and completeness of description within the confines of a problem; it is a definitive
answer. Interpretation involves judgment, experience, familiarity with a problem and a host of other
factors. It is subjective, but it is based on a set of skills which can be developed in time. Most of us use
the seven image characteristics subconsciously. To improve our image interpretation skills, however, we
should consciously examine the image characteristics associated with the themes we are interested in
interpreting, and ensure that the discernible characteristics are leading to the same interpretation.
Again, this requires taking the imagery into the field.

8
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

IMAGE INTERPRETATION TOOLS

Currently, the most efficient image interpretation tool is the human eye-brain system of a skilled image
interpreter. However, manual interpretation is subjective and lacks rigor. While still based on image
interpretation, Image Keys, Image Sampler, and Digital Cover Interpreter are examples of tools that
help provide consistent and objective results.

Image Keys
Image keys can take many forms but, essentially, they are a set of guidelines designed to aid the
interpreter in correctly applying the classification scheme during the interpretation. Image keys also help
to ensure that consistency is maintained throughout the interpretation process. Keys may also be
valuable teaching aids for persons who lack field experience, yet who must be quickly trained to identify
features on a given type, scale, and quality of imagery.

One of the simplest and most effective image key consists of saving a copy of previously interpreted
imagery (that has been carefully ground-checked for accuracy) as a "training set" or "reference set" to
which unknown areas of the imagery may be compared. While being simple and effective, this type of
image key may have limited utility with different types of imagery, different magnification of inspection,
or with very complex keys.

It can be very helpful to also include written descriptions of the features of interest which can inform the
interpreter of what to look for in the imagery. It may also include tips such as “this class is sometimes
confused with class X, however, they can usually be differentiated by …”. Adding additional habitat
information can also help the interpreter—for example “species X typically occurs between 4700 feet
(1430 meters) and 6200 feet (1890 meters) in elevation—slightly higher on south facing slopes and
lower on north facing. This species is often co-dominant with species Y.”

It is valuable to make keys as a learning exercise. They are great training aids for specific projects when
several interpreters must be trained to reliably and consistently identify a specific and restricted set of
tree/stand conditions. One of the best ways of learning to differentiate the details of tree species at any
scale is to make a key to the species imaged. Making such a key requires several field identified
examples of each species. Using such a key for training purposes requires reference to the same set of
field identified examples. It is advisable to collect a personal "portfolio" of examples from a range of
conditions. Field work is crucial for developing image interpretation skills and for training others.

Image Sampler and Digital Cover Interpreter

These are example digital tools (ArcMap Add-ins) that can help provide consistency and objectivity to
your ocular interpretations.

Image Sampler creates sample points within a user-specified polygon and provides a toolbar interface
for attributing those sample points. These sample points can then be used to generate statistical

9
 IMAGE INTERPRETATION

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

estimates of the cover types within the polygon(s).

Cover Interpreter was designed to help users

adhere to established mapping standards and
produce existing vegetation GIS layers. It allows the
user to assign cover attributes (canopy cover and
size class) to polygons by comparing the underlying
image to a series of transparent digital image chips
that can be moved over the image.

These digital tools, while consistent and objective,

are slower than ocular interpretations. Perhaps the
best way to use them in a large project is more at
the start of a project to grow confidence and
interpretation skill and sparingly thereafter for
occasional recalibration of our eye-brain system.

Image Sampler Tool—dot grid can be systematic (as

shown) or random. The number of dots within the
polygon is user-defined. After each dot is attributed,
proportions can easily be calculated for the polygon.

Image Sampler and Cover Interpreter are

available for free download from:
[Link]

Cover Interpreter Tool—used to quickly assess per-

cent canopy cover.

10
 APPENDIX 1—SUGGESTED IMAGE
INTERPRETATION WORKFLOW
RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

The following are suggested guidelines for interpretation of vertical high-resolution* orthocorrected
digital imagery. As you prepare to engage in image interpretation, these reference guidelines will help
build skills and maintain consistency throughout the interpretation process. It is assumed that you have
adequate ground-based knowledge about the resource you wish to interpret from the imagery and that
your objective is a continuous (wall-to-wall) polygon-based classification.

SUGGESTED STEPS
1. Define and be very familiar with the classification scheme (see Appendix 2)
2. Prepare the imagery
a. Imagery should be accurately orthocorrected and displayed in your corporate GIS/image
processing software
3. Segment the image, using heads-up digitizing, existing polygon data, or using image segmentation
software, into appropriate mapping polygons. You will want to determine a minimum mapping unit
(MMU) and ensure that few, if any, polygons are smaller than the MMU—unless the segmentation has
been deliberately designed to over-segment the landscape. If the landscape has been over-
segmented, many segments will need to be merged later.
4. Generate Field Maps with the segments and background imagery.
a. Create a field map for each segment/polygon you will visit in the field.
i. The maps should contain the outlined (aka boundary) polygon, the high-resolution
imagery, a north arrow, scale bar, grid lines (spaced at an appropriate interval to be
meaningful), coordinate for the center point of the polygon, and any other notes or
information that will make the map more meaningful in the field.
b. Create an overview map(s) of all the polygons.
i. These map(s) should contain an overview of all the polygons that will be visited in the
field. If you will be visiting polygons that are fairly dispersed in your project area, it is
appropriate to create many overview maps to ensure you capture all the field
polygons . The overview maps should also contain the same items as in the field
maps—north arrow, scale bar, grid line, and relevant coordinates.
c. If you have a GPS unit, upload reference locations to help you navigate to the polygons.
4. Take the imagery with overlaying segment boundaries into the field (hardcopy prints or digital
mobile device). The purpose of this visit is to start the familiarization process and, to test the
classification scheme and image capabilities.
a. Select several polygons with different characteristics to visit:
i. If you are already familiar with the area, select the polygons based on known

* High spaƟal resoluƟon imagery in this case means 1 meter ground sample distance (pixel size) or finer.
11
 APPENDIX 1—SUGGESTED IMAGE
INTERPRETATION WORKFLOW
RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

differences in your classification scheme (e.g., big trees, small trees or evergreen,
deciduous). Try to select polygons that represent all of the classes in your
classification scheme to visit (thus the complexity of your classification scheme will
drive the number of polygons that you should visit).
ii. If you are unfamiliar with the area, select the polygons with differences in the image
characteristics that you suspect portray differences in the feature(s) of interest.
Eventually, you should visit polygons that represent all of the classes in your
classification scheme.
b. For each visited polygon:
i. Try to classify it based solely on ground observations.
ii. Carefully inspect the same polygons in the imagery while you’re within each polygon.
Try to build the bridge from what you’ve seed on the ground to what you can see on
the image. Consider making notes on your field form (if you’ve developed a field form
yet); otherwise write notes on the field maps for that polygon.
c. After visiting the selected polygons on the ground, classify the same polygons based solely on
your interpretation of the imagery.
d. Check your image interpreted classifications against your ground-based classifications.
i. If results indicate you can successfully interpret the imagery, proceed to the next step
in this process. However, if results indicate you cannot interpret the imagery
successfully, is the problem a classification-imagery mismatch (i.e., there are no
discernible image characteristics that permit interpretation of the feature(s) of
interest)? If you have a hierarchical classification, can you collapse it a level (less
detailed) and have success (and an acceptable product)? Or, would more image
interpretation practice solve the problem sufficiently?
e. Now interpret and classify polygons on the imagery that you have not visited on the ground—
try to include all the classes in your classification scheme.
f. Visit those polygons on the ground and check your image interpreted classification.
i. If results indicate you can successfully interpret the imagery, proceed to the next step
in this process. However, if results indicate you cannot interpret the imagery
successfully, is the problem a classification-imagery mismatch (i.e., there are no
discernible image characteristics that permit interpretation of the feature(s) of
interest)? If you have a hierarchical classification, can you collapse it a level (less
detailed) and have success (and an acceptable product)? Or, would more image
interpretation practice solve the problem sufficiently? Make sure the issue is addressed
before proceeding further with this project!

12
 APPENDIX 1—SUGGESTED IMAGE
INTERPRETATION WORKFLOW
RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

5. With success to this point, you should develop image keys, field forms, and image training
sets.
6. Classify polygons on the digital imagery.
a. Attribute easy features first (e.g., water vs. forest) and go from the known to the unknown
b. After initial broad classification, further subdivide using more subtle features
 For timber mapping: type density (canopy cover) first, species (dominance class) next,
and then size class
 Use both high and low zoom magnification—different magnifications can accentuate
different image attributes
c. Undoubtedly you will encounter polygons that need to be split or joined—make sure you are
familiar with these digitization techniques. If you are splitting a polygon, make sure that all
parts exceed the MMU size.
7. Check your work and recalibrate your interpretations frequently. It is common to have
“classification drift” which means as the project progresses the way you interpret features changes.
Classification drift occurs during the course of an individual day (usually in a bad way due to fatigue)
and also over the course of weeks and months as you become more skilled.
a. Frequently refer to your image keys and training sets.
b. Occasionally use a tool such as Image Sampler, after you have classified a polygon by eye, to
test your determinations. Note: it’s preferable for efficiency to have the confidence and skill
to perform the classifications manually (by eye) rather than using Image Sampler. Image
Sampler is more consistent and unbiased than a human interpreter but it is slower. Its best
use in this workflow is as a recalibration tool for your manual interpretations.

13
 APPENDIX 2—THE CLASSIFICATION
SCHEME

RSAC/RSEAT—Remote Sensing Evaluation, Application & Training [Link]

Classification schemes categorize and label the variation in land cover that you want to capture. A
classification scheme is analogous to the legend of the final map you wish to produce. A workable
classification scheme is the first step in a successful image-interpretation project and should not be
taken lightly.

High-resolution imagery has a tremendous amount of information—often too much information to

process efficiently. We need to extract and summarize the information to make it manageable. Thus,
image interpretation projects focus on a particular type of information—themes—to extract from the
imagery. For example, a given project may focus on gathering information about varying levels of
vegetation information despite the imagery having a nearly unlimited number of other potential themes -
other examples of themes might include forest health, transportation features, water features, cultural
surveys, and many more.

Following are five rules and three suggestions to keep in mind about a workable classification scheme.
First the rules:

 The classification scheme must meet the information needs of the intended user.
 The classification scheme must be based on what can be interpreted from the imagery (note
that this may conflict with the previous requirement in some cases).
 Continuous classification schemes must be exhaustive.
 The classification scheme must be mutually exclusive.
 The classification scheme must be composed of labels and rules for labeling.

In addition to the above five requirements, the classification scheme should:

 Be hierarchical. A hierarchical (or nested) classification scheme will provide for more flexibility
and provide better support to multiple (and diverse) users.
 Be based on measurable land cover (not the specific land use) characteristics.
 Avoid the more subjective interpretive classes such as "Old Growth" or "Suitable Habitat" -
these items in the classification scheme should be derived by definitions based on measurable
feature characteristics.

Ultimately, the complexity of a classification scheme will affect the project accuracy and the project cost.
The more complex the scheme, the more expensive and the less accurate the final product will be. A
good and workable classification scheme is not a guarantee of success, but a poor and unworkable
scheme is a guarantee of failure. Give it the time and attention it deserves for your project.

14