Digital Image Processing Part 2

1
Digital Image ProcessingPart 2
2
Image Classification
(ccrs.nrcan)

• The objective is to automatically categorize all
  pixels into land cover classes or themes.
• Different feature types manifest different
  combinations of Digital Numbers (DNs) based on
  their inherent spectral reflectance and emittance
  properties.

3
Image ClassificationExample
SPOT before

• Example
• Color in map and Landcover type
• (black) Clear water.
• (green) Dense forest with closed canopy.
• (yellow) Shrubs, less dense forest.
• (orange) Grass.
• (cyan) Bare soil, built-up areas.
• (blue) Turbid water, bare soil, built-up areas.
• (red) Bare soil, built-up areas.
• (white) Bare soil, built-up areas.

SPOT after thematic mapping
(Virtual Science Centre)
4
Image ClassificationExample
Mean pixel values from previous data
(Virtual Science Centre)
5
Image Classification

• Three types of classification procedures
• Supervised
• Analyst supervises pixel categorization process
  using training areas to compile interpretation
  key (training step followed by classification
  step).
• Unsupervised
• Similar to supervised, but two separate steps
  differing in that data is classified by
  aggregating into natural spectral groupings or
  clusters, then analyst determines identity by
  comparing to reference data.
• Hybrid
• Mixture.

6
Image Classification Supervised Classification

• Much more accurate for mapping classes, but
  depends heavily on skills of image specialist.
• Specialist must recognize classes in a scene from
  prior knowledge, with
• Experience with region
• Experience with thematic maps
• On-site visits
• Familiarity allows specialist to choose and set
  up discrete classes.
• Then supervision of selection of classes and
  assignment of category names.

7
Image Classification Supervised Classification

• Next location of training sites on the image to
  identify the classes.
• Training sites are areas representing each known
  land cover category that appear fairly
  homogeneous on image.
• Similarity in tone or color within shapes
  delineating category.
• Specialists locate and mark them with polygonal
  boundaries on image display.
• Mean values and variances of DNs for each band
  used to classify them calculated from all pixels
  in the site.
• Result is a spectral signature for that class.

8
Image Classification Supervised Classification
(Concluded)

• Classification now proceeds by statistical
  processing
• Every pixel is now compared with various
  signatures and assigned to the class whose
  signature comes closest.
• A few pixels in scene may not match and will
  remain unclassified.
• May belong to a class not recognized or defined.

9
Image Classification The Training Stage

• Actual classification of multispectral image data
  is highly automated.
• Training data needed to support automation.
• Must be representative and complete.
• Training statistics for all spectral classes
  constituting each information class.
• Information class could be grass or water
• If interested in both clear and turbid water,
  minimum of two spectral classes would be required
  to adequately train on this feature.
• One or more of the following types of analyses
  are typically involved in the training set
  refinement process.
• Graphical representation of spectral response
  pattern.
• Quantitative expressions of category separation.
• Self-classification of training set data.
• Interactive preliminary classification.
• Representative sub-scene classification.

10
Image Classification The Training Stage
(Continued)

• Analyst identifies homogeneous representative
  samples of different surface cover types.
• Referred to as training areas.
• Analyst is supervising the categorization of a
  set of specific classes.
• Numerical information in all spectral bands for
  the pixels comprising these areas are used to
  train the computer to recognize spectrally
  similar areas for each class.
• In supervised classification, information
  classes are first identified to determine
  spectral classes which represent them.

(ccrs.nrcan)
11
Image Classification The Training Stage
(Concluded)

• Morro Bay Training Site
• True color bands (bands 1,2,3)
• Site colors for display convenience

(rst)

• Plots for 8 general categories
• Multiple training sites
• Greatest separability in band 5

(rst)
12
Image Classification The Classification Stage

• There are four classifiers
• Measurements on scatter diagram
• Minimum distance to mean classifier/centroid
  classifier
• Parallelpiped classifier
• Gaussian maximum likelihood classifier

13
Image Classification Minimum-Distance-to-Means
Classifier

• Supervised classification
• DNs in multidimensional band space.
• Unknown pixel placed in class closest to mean
  vector in band space.

• Morro Bay, Minimum Distance Classification.
• All seven TM bands including thermal 16 gray
  levels with arbitrary color assigned.

(rst)
14
Image Classification Maximum Likelihood
PDF

• Assumption of normality
• Distribution of category described by mean
  vector and covariance matrix.
• Can compute statistical probability of given
  pixel value being a member of particular class.
• Vertical axis associated with probability of
  pixel value being a member of one of the classes.
• Resulting bell-shaped surfaces called probability
  density functions.
• One such function for each spectral category.
• Compute probability of pixel belonging to each
  category.

Band 4 digital number
Band 3 digital number
(rst)
15
Image Classification Maximum Likelihood
(Lillisand)
16
Image Classification Unsupervised
Classification Methodology

• Reverses the supervised classification process.
• Spectral classes are grouped first, based on the
  datas numerical information, and they are then
  matched by the analyst to information classes.
• Clustering algorithms are used to determine the
  natural (statistical) groupings of structures in
  the data.
• Unsupervised classification is not completely
  without human intervention, but it does not start
  with a pre-determined set of classes.

(ccrs.nrcan)
17
Image Classification Mixed Pixels

• Individual areas consisting of different features
  or classes.
• Below resolution of sensor
• Treat as more or less homogeneous
• Resulting spectral content is composite of
  weighted average.

(rst)
18
Image Classification Fuzzy Classification

• Attempts to handle mixed-pixel problem by
  fuzzy-set concept.
• Similar to K-means unsupervised classification,
  but no hard boundaries between classes in
  spectral measurement space.
• Fuzzy boundaries instead of unknown measurement
  vector being assigned solely to a single class.
• Membership grade values assigned that describe
  how close a pixel measurement is to the means of
  all classes.
• Another approach is fuzzy supervised
  classification.
• Similar to application of maximum likelihood
  classification, but fuzzy mean vectors and
  covariance matrices are developed from
  statistically weighted training data.
• Combinations of pure and mixed training sites may
  be used
• Known mixtures of various feature types define
  fuzzy training class weights.
• Classified pixel is assigned a membership grade
  with respect to its membership in each
  information class.

19
Image Classification The Output Stage

• Utility of any image classification ultimately
  dependent on production of output products that
  effectively convey interpreted information to its
  end user.
• Interface boundaries amongst destinations become
  blurred
• Remote sensing
• Computer graphics
• Digital cartography
• GIS Management
• Critical component is ability to provide graphics
  that convey results of analysis to people who
  make decisions about resources.

20
Image Classification Post-classification
Smoothing

• Classification data often have a salt and pepper
  appearance.
• Spectral variability encountered by a classifier
  when applied on a pixel-by-pixel basis.
• Several pixels scattered throughout corn field
  may be classified as soybeans.
• Often desirable to smooth classified output to
  show only the dominant (presumed correct)
  classification.
• One might consider low pass filter, but there is
  a problem.
• Output of image classification is array of pixel
  locations containing labels such as land cover,
  not quantities.
• Post-classification smoothing algorithms must
  operate on basis of logic rather than simple
  arithmetic operations.

21
Image Classification Classification Accuracy
Assessment

• Four tests
• Field checks of selected points
• Estimate of agreement between classification maps
  and reference maps.
• Statistical analysis of numerical data using
  tests such as correlation coefficients.
• Overlay with re-scaled aerial photo until field
  patterns approximately match.

(rst)
22
Image Classification Accuracy Assessment

• Two remote sensing bands better than one for
  producing largest improvement in in accuracy.
• Information gained after four bands not as great
  accuracy increase flattens or increases very
  slowly.
• Additional bands such as TM bands 5 and 7 can be
  helpful in identifying rocks.
• More than one band better for identifying crop
  types.

23
Data Merging and Geographic Information Systems
(GIS) Integration

• Relating information from
• different sources
• Data capture
• Data integration
• Projection and registration
• Data structures
• Data modeling

(rst)
24
Data merging and GIS integration GIS Data
Integration
(rst)
Geographical Information Systems makes it
possible to link, or integrate, information that
is difficult to associate.
25
Hyperspectral Image Analysis

• GIS relates information from different sources
• Data capture
• Data integration
• Projection and registration
• Data structures
• Data modeling
• Multisensor image merging often results in a
  composite image product that offers greater
  interpretability
• Can merge multispectral sensor and radar image
  data
• Spectral resolution of multispectral scanner data
• Radiometric and sidelighting characteristics of
  radar data.

26
Hyperspectral Image Analysis
http//satjournal.tcom.ohiou.edu/
27
Biophysical Modeling

• Biophysical modeling is a combination of physical
  modeling and empirical modeling.
• An example might be remote sensing data with
  ground measurement controls.
• Consider mapping earthquake faults
• There might be satellite imaging in several
  bands.
• Included with ground inspection of rock types,
  fault movement, and damage.

28
Image Transmission and Compression

• Large amounts of data are used to represent a
  typical image.
• Because technology permits ever-increasing image
  resolution and increasing numbers of spectral
  bands, there is a consequent need to limit the
  resulting data volume for speedier transmission.
• The amount of storage media needed is enormous.
• One possible approach to decreasing the necessary
  amount of storage is to work with compressed
  image data.

Data Compression and image reconstruction
(http//www.eng.iastate.edu/ee528/sonkamaterial/c
hapter_13_image_data_compressio.htm. )
29
Image Transmission and Compression (Continued)

• Data compression methods can be divided into two
  principal groups
• Information preserving compression permit
  error-free data reconstruction (lossless
  compression).
• Compression methods with loss of information do
  not preserve the information completely (lossy
  compression).
• In image processing, a faithful reconstruction is
  often not necessary in practice and then the
  requirements are weaker, but the image data
  compression must not cause significant changes in
  an image.

30
Image Transmission and Compression (Continued)

• Discrete cosine image compression.
• Reconstructed image
• Differences between pixel values in original and
  to reconstructed image

31
Supplemental References

• Remote Sensing Tutorial, http//rst.gfsc.nasa.gov
• Image Interpretation and Analysis,
  http//www.ccrs.nrcan.gc.ca/ccrs/eduref/tutorial/c
  hap4/c4p6e.html
• Geographic Information Systems,
  http//www.usgs.gov/research/gis/application
• Color Representations, http//www.geo.utep.edu/pub
  /keller/color.html