Applied Analysis
Inc.
630 Boston Road
Suite 201
Billerica, MA 01821
USA
PH: 978-663-6828
FAX: 978-663-6389
|
|
|
|
 |
|
|
Frequently
Asked Questions
| FAQs:
General |
|
1. What is IMAGINE Subpixel Classifier?
|
| IMAGINE
Subpixel Classifier is a supervised, non-parametric spectral classifier
that performs subpixel detection and quantification of a specified
material of interest. The process allows you to develop material signatures
and apply them to classify image pixels. It reports the material pixel
fraction occupied by the material of interest and may be used for materials
covering as low as 20% of a pixel. Its unique image normalization process
allows you to apply signatures developed in one scene to other scenes
from the same sensor. IMAGINE Subpixel Classifier is a fully integrated,
add-on application module to ERDAS IMAGINE.
Because it addresses the "mixed
pixel problem," IMAGINE Subpixel Classifier successfully identifies
a specific material when other materials are also present in a pixel.
It discriminates between spectrally similar materials, such as individual
plant species, specific water types, or distinctive building materials.
And it allows you to develop spectral signatures that are scene-to-scene
transferable.
In its early versions, IMAGINE Subpixel
Classifier was called AASAP (Applied Analysis Spectral Analytical Process),
a name that occasionally still appears in remote sensing literature. |
|
|
|
3. What does "subpixel" mean?
|
| Most
image pixels are mixed pixels, i.e., they contain more than one material
or feature. Any material that does not fully occupy a pixel is considered
a subpixel component of that pixel. Subpixel classification involves
the detection and identification of materials that cover less than
the full extent of a pixel.
IMAGINE Subpixel
Classifier provides the capability to detect and quantify these subpixel
occurrences. It enables the detection of small objects and more accurate
classification of large area features that are not homogeneous.
|
|
|
|
4. How does IMAGINE Subpixel Classifier differ from traditional
classifiers?
|
|
The primary
difference between IMAGINE Subpixel Classifier and traditional classifiers
is the way in which it derives a signature from the training set and
then applies it during classification. Traditional classifiers typically
form a signature by averaging the spectra of all training set pixels
for a given feature. The resulting signature contains the contributions
of all materials present in the training set pixels. This signature
is then matched against whole-pixel spectra found in the image data.
In contrast, IMAGINE Subpixel Classifier
derives a signature for the spectral component that is common to the
training set pixels following background removal. This is normally
a pure spectrum of the material of interest (MOI). Since materials
can vary slightly in their spectral appearance, IMAGINE Subpixel Classifier
accommodates this variability within the signature. The IMAGINE Subpixel
Classifier signature is therefore "purer" for a specific material and
can more accurately detect the MOI. During classification, the process
subtracts representative background spectra to find the best fractional
match between the pure signature spectrum and candidate residual spectra. |
|
|
|
5. Do I use IMAGINE Subpixel Classifier in place of a
traditional classifier?
|
| IMAGINE
Subpixel Classifier and traditional classifiers perform
best under different conditions. IMAGINE Subpixel Classifier should work
better to discriminate among species of vegetation, distinctive building
materials, or specific types of rock or soil. You would use it to find
a specific material, even when it covers less than a pixel.
You may prefer
a traditional classifier when the MOI is composed of a spectrally varied
range of materials that must be included as a single classification unit.
For example, a forest that contains a large number of spectrally distinct
materials and spans multiple pixels in size may be classified better
using a minimum distance classifier. IMAGINE Subpixel Classifier can
compliment a traditional classifier by identifying subpixel occurrences
of specific species of vegetation within that forest.
When
deciding to use IMAGINE Subpixel Classifier, recall that it identifies a
single material – the MOI, whereas a traditional classifier will
classify many materials or features occurring with a scene.
|
|
|
|
6. How is IMAGINE Subpixel Classifier different from
other subpixel classification techniques?
|
|
Various methods
of spectral mixture analysis have been developed to improve the
classification of mixed pixels, and to detect and identify subpixel
components and their proportions. Most of the techniques employ a linear
mixing approach.
Linear mixing
assumes that radiation reflected from one material in the pixel field
of view has not transmitted through or reflected from any other ground
materials before emerging toward the sensor. Under these conditions,
the light reflected from the individual ground components linearly
combine in proportion to the abundance of each ground component.
The resultant
pixel spectrum can in turn be linearly unmixed to reveal the individual
ground components, if the spectra of the individual components are
known. One of the most commonly used linear unmixing techniques is
the Linear Mixing Model (LMM).
While IMAGINE
Subpixel Classifier employs a linear unmixing technique, it differs
significantly from LMM models in both approach and discrimination performance
characteristics. The LMM models
each pixel spectrum as a linear sum of a common set of image end-member
spectra. The total number of end-members is n-1 or less, where n is
the number of spectral bands. For Landsat TM, the maximum number of
end-members is five.
The end-member
spectra include a shade spectrum, a set of "background" spectra, and
a set of "residual" spectra. The background end-members are assumed
to be in every pixel, while the residual end-members are assumed to
be in only some of the pixels. The end-member spectra are selected to
be as different as possible from one another (ideally they should be
spectrally orthogonal.)
The output
is a set of fraction planes, one for each end-member spectrum, which
gives the relative amounts of each end-member in each pixel. A residual
plane is also produced that provides the root-mean-square error of
the fit for each pixel.
IMAGINE Subpixel
Classifier, in contrast, models each pixel as a linear combination
of only two components, the MOI and the background. The MOI is assumed
to be in every pixel, while the background component is assumed to
be unique to each pixel.
The user provides
the spectrum (signature) for the MOI. The background spectra (44 candidate
backgrounds) are derived autonomously by the process. There is no requirement
for spectral orthogonality, i.e., the background can be spectrally
similar to the MOI. The output is a single fraction plane, listing
the amount of the MOI in each pixel. |
|
|
|
7. What types of materials can I classify with IMAGINE
Subpixel Classifier?
|
| In principle,
IMAGINE Subpixel Classifier can be used to classify any material that
has a distinct spectral signature relative to other materials in a
scene. IMAGINE Subpixel Classifier has been most thoroughly evaluated
for vegetation classification applications in forestry, agriculture,
and wetland inventory as well as for man-made objects, such as construction
materials. IMAGINE Subpixel Classifier has also been used in defining
roads and waterways. |
|
|
|
8. How accurate is IMAGINE Subpixel Classifier?
|
|
Classification accuracy depends on many factors. Some of
the most important are:
-
Number of spectral bands in the imagery.
Discrimination capability increases with the number of bands. Smaller
pixel fractions can be detected with more bands. The 20% threshold
used by the software is based on 6-band data.
-
Target/background contrast.
-
Signature quality. Ground truth information
helps in developing and assessing signature quality.
-
Image quality, including band-to-band registration,
calibration, and resampling (nearest neighbor preferred)
Two
projects involving the subpixel classification of
wetland tree species Cypress and Tupelo and of invasive
forest tree species Loblolly pine included extensive
field checking for classification refinement and accuracy
assessment. The classification accuracy for these
materials was 85-95%. Classification of pixels outside
the training set area was greatly improved by IMAGINE
Subpixel Classifier in comparison to traditional classifiers.
In a separate
Quantitative
Evaluation study designed to assess the accuracy of IMAGINE Subpixel
Classifier, hundreds of man-made panels of various known sizes were deployed
and imaged. The approximate amount of panel in each pixel was measured.
When compared to the Material Pixel Fraction (the reported amount of
material in each pixel) reported by IMAGINE Subpixel Classifier, a high
correlation was measured. Also, IMAGINE Subpixel Classifier outperformed
a maximum likelihood classifier in detecting these panels. It detected
190% more of the panel containing pixels with a lower error rate and
reported the amount of panel in each pixel classified.
Depending on
the spectral separability of a material and its size range, lower accuracies
may be achieved. The same may be true for scene-to-scene applications
where the material may vary between two locations. The IMAGINE Subpixel
Classifier Signature Refinement process is designed to help improve
accuracy in scene-to-scene applications.
In general,
you should keep in mind that there will be a trade-off between classification
accuracy and false alarm rate, just as with any detection process.
The classification process within IMAGINE Subpixel Classifier provides
a tolerance parameter that allows you to increase or decrease the spectral
filter tolerance. By adjusting this tolerance parameter, you can achieve
different levels of MOI detection probability versus false alarm rate.
Classification accuracy can be quantified in terms of a Receiver Operating
Curve (ROC) in which probability of detection (or detection rate) is
plotted against false alarm rate with different values on the curve
resulting from different tolerance parameter values. The optimum tolerance
value for your application can be determined using this methodology.
|
|
|
|
9. What data sources can I use with IMAGINE Subpixel
Classifier?
|
| IMAGINE
Subpixel Classifier works on any multispectral data source, including
airborne or satellite, with three or more spatially registered bands.
The data must be in either 8-bit or 16-bit format. Landsat
Thematic Mapper (TM), SPOT XS and Xi, and IKONOS multispectral imagery
have been most widely used because of data availability. It will also
work with data from other high-resolution commercial sensors such as
QuickBird, Airborne sources, and OrbView-3. When ordering Landsat TM
data for use with IMAGINE Subpixel Classifier, the preferred data format
is 30m resampling, spacecraft path orientation, radiometrically-corrected,
geometrically uncorrected, with nearest-neighbor resampling. When ordering
SPOT data, the preferred format is
Level 1A.
IMAGINE Subpixel Classifier will also work with most hyperspectral
data sources. In this case, the software will automatically uniformly
downsample the spectral data to a multispectral set during certain
operations. Better classification results may be achieved by first
downsampling the data to a multispectral data set using a band set
optimized for your application. |
|
|
|
10. Does IMAGINE Subpixel Classifier provide value
when working with high spatial resolution data sources?
|
|
Yes. There is considerable utility
in using IMAGINE Subpixel Classifier to exploit high-resolution data
such as QuickBird and IKONOS. Regardless of the resolution, the mixed
pixel problem will always exist. In addition, the higher
spatial resolution allows for signature derivation of spatially smaller
targets. Furthermore, certain targets exhibit variations at a higher
resolution that would not be observed at a coarser resolution.
Some issues to consider when using
high-resolution data are that generally there may be a lower spectral
resolution due to the decrease in the number of bands. In
addition, IKONOS and QuickBird lack the SWIR band useful for vegetation
discrimination. Also, certain targets can exhibit variations
at higher resolution that would be blended together in low resolution
data. This may mean that multiple signatures are required to properly
classify all occurrences of and the full extent of the material of
interest. |
|
|
|
11. What do I need to know to use IMAGINE Subpixel
Classifier?
|
| A typical
ERDAS IMAGINE user with prior experience using traditional supervised
multispectral classifiers can be up and running with IMAGINE Subpixel
Classifier in less than a day. You need only learn the specific signature
derivation and image classification techniques for IMAGINE Subpixel
Classifier. Familiar IMAGINE tools are used for importing data, viewing
images, creating training sets, and overlaying results. As with
any classifier, ground truth is useful in selecting training set pixels
and for assessing accuracy. |
|
|
|
|
|
|
|
AAI's Alaskan Oil Spill Assessment Using Satellite Imagery. (more...)
|
|
Demo Web-based mapping using SVG (more)
|
|
