Image Classification

Download Report

Transcript Image Classification

Image Classification
3:37 AM
Image Classification
Automatically categorize all pixels in an image into
land use/cover classes or themes.
A process of thematic information extraction
A process of pattern recognition
3:37 AM
Image Classification
• The process of arranging raw data DNs into information classes.
• Two Basic Types
• Supervised
• Unsupervised
Data
Information
Raw Imagery
3:37 AM
Extracted Information
Supervised Classification
•The image analyst “supervises’ the pixel categorization
process specifying, to the computer algorithm,
numerical descriptors of the various land cover types
present in a scene.
•Representative sample sites of known cover type,
called training areas, are used to compile a numerical
“interpretation key” that describes the spectral attributes
for feature type of interest.
3:37 AM
Unsupervised Classification
In an unsupervised classification, the computer groups
pixels with similar spectral characteristics into unique
clusters according to some statistically determined
criteria.
The analyst then labels and combines the spectral
clusters into information classes.
3:37 AM
Land-use and Land-cover Classification Schemes
Land cover refers to the type of material present on the landscape
(e.g., water, sand, crops, forest, wetland).
Land use refers to what people do on the land surface (e.g.,
agriculture, commerce, settlement).
3:37 AM
Land-use and Land-cover Classification Schemes
A classification scheme contains taxonomically
correct definitions of classes of information that are
organized according to logical criteria.
The classes in the classification system should
normally be:
• mutually exclusive,
• exhaustive, and
• hierarchical.
3:37 AM
Land-use and Land-cover Classification Schemes
•Mutually exclusive means that there is no taxonomic
overlap of any classes (i.e., deciduous forest and
evergreen forest are distinct classes).
•Exhaustive means that all land-cover classes present in
the landscape are accounted for and none have been
omitted.
* Hierarchical means that sublevel classes (e.g., singlefamily residential, multiple-family residential) may be
hierarchically combined into a higher- level category
(e.g., residential) that makes sense.
3:37 AM
Land-use and Land-cover Classification Schemes
It is also important for the analyst to realize that there
is a fundamental difference between information
classes and spectral classes.
* Information classes are those that human beings
define.
* Spectral classes are those that are inherent in the
remote sensor data and must be identified and then
labeled by the analyst.
3:37 AM
U.S. Geological Survey LandUse/Land-Cover Classification
System for Use with Remote Sensor
Data
3:37 AM
Feature Space Scatter plots
• Compares two image
bands in feature
space
• Basically two
histograms displayed
on two perpendicular
axes.
The brighter a particular point is in
the display, the more pixels within
the scene having that unique
combination of band values.
3:37 AM
Some Feature Space Concepts
3:37 AM
Pixel Position (X,Y)
255
64, 191
Band Y
191
191, 127
127
64, 64
64
0
3:37 AM
64
127
Band X
191
255
Scatterplot: High Correlation
Image with 5 pixels
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Highly Correlated
3:37 AM
Low Correlation
Image with 5 pixels
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Low Correlation
3:37 AM
Well Defined Clusters
Image with 30 pixels and 5 clusters
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Not So Well Defined Clusters
Image with 30 pixels and 5 clusters
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Poorly Defined Clusters:
Some Class Confusion
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Very Poorly Defined Clusters:
Total Class Confusion
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Calculating Cluster Mean
Mean for Red-Dots Cluster
10
Cluster Mean for X position 
7.5
Cluster Mean for Y position 
x
n
 yi
n
Band Y
X Mean: (2.5+2.5+5+ 5+10) / 5 = 5
5
Y Mean: (2.5+ 2.5+ 5+10+10) / 5 = 6
Cluster Mean = 5,6
2.5
0
3:37 AM
2.5
5
7.5
10
i
Calculating Cluster Variance
Var(x) 
2


x

x
 i
 y
Var(y) 
10
n
 y
2
i
n
Cluster Mean For x = 5
Band Y
7.5
Cluster Mean For y = 6
5
Variance(X) for Cluster = 7.5
Variance(Y) for Cluster = 8
2.5
0
2.5
5
7.5
10
For Var(x) = [(2.5 - 5)2 + (2.5 - 5)2 + (5 - 5)2 + (5 - 5)2 + (10 - 5)2 ]/5 = (6.25 + 6.25 + 0 + 0 + 25)/5 = 37.5 / 5 = 7.5
For Var(y) = [(2.5 - 6)2 + (2.5 - 6)2 + (5 - 6)2 + (10 - 6)2 + (10 - 6)2] /5 =(3.5 + 3.5 + 1 + 16 + 16/5 = 40/5 = 8
3:37 AM
Calculating Cluster Standard Deviation
1 Standard Deviation
Standard Deviation in X Direction 
10
Standard Deviation in Y Direction 
2


x

x
 i
n
 y
 y
2
i
n
Band Y
7.5
Variance(X) for Cluster = 7.5
Variance(Y) for Cluster = 8
5
2.5
1 Standard Deviation (X)  7.5  2.739
1 Standard Deviation (Y)  8  2.828
0
2.5
5
Band X
3:37 AM
7.5
10
Calculating Distance in 2d Space
100
Distance  (rise ) 2  (run) 2
Band Y
75
50
25
2500  5625  8125  90.139
0
25
3:37 AM
Band X
50
75
100
Parallelepiped View: Standard Deviation
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Class Ellipse View: Standard Deviation
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Parametric vs. Non-Parametric
Signatures
• A parametric signature is based on a statistical analysis of
the pixels that are in the training site or a cluster.
– A parametric decision rule uses statistical analysis to assign
pixels in an image to a particular class. When a given pixel
meets the parameters of the decision rule for a given class the
pixel is assigned to that class.
• A non-parametric signature is not based on statistics, but on
discrete objects within feature space.
– With a non-parametric decision rule, if a pixel is located
within the boundary of the non-parametric signature in feature
space then that pixel will be assigned to the category
represented by the signature.
3:37 AM
Parametric vs. Non-Parametric
Signatures
Parametric: Statistically Defined
Non-Parametric: User Defined
255
255
TM 4
TM 4
0
255
TM 3
3:37 AM
0
255
TM 3
User Defined Signature (Non-Parametric)
255
Band Y
191
127
64
0
3:37 AM
64
127
Band X
191
255
Supervised Classification
• Supervised: A method in which the analyst spatially defines training sites
representative of each desired class (category). The analyst then "trains" the
computer software to recognize varying spectral values in two or more
spectral bands associated with those training sites. This is called signature
definition. After signatures for each category have been defined, the computer
then uses those signatures to classify the remaining pixels in the study area.
Data
Information
Raw Imagery
3:37 AM
Training Data
Collection
Signature Creation
Training Sites
Extracted Information
1. Define Classification Scheme
-Example: Land Use/Cover Classification Scheme, Atlanta Georgia
Metropolitan Area. Classification of Landsat TM images.
Code
s
3:37 AM
Land use/cover
classes
Description
1
High-density
urban
Central business districts, multi-family dwellings,
commercial and industrial facilities, high impervious
surface areas of institutional facilities, large transportation
facilities, e.g. airports, multilane interstate/state highways
2
Low-density
urban
Single family residential areas, urban recreational areas,
cemeteries, playing fields, campus-like institutions, parks,
schools, local roads
3
Bare land
Areas with sparse vegetation (less than 20%), forest clearcut, fallowed cropland, quarries, strip mines, rock outcrop,
sand beach along rivers and lakes
4
Cropland or
grassland
Row crop agriculture, orchids, vineyards, horticultural
businesses, pastures, non-tilted grasses, golf courses
5
Forest
Evergreen forest, deciduous forest, and mixed forest
6
Water
Rivers, streams, lakes, and reservoirs
2. Collect Training Data.
•GPS Field Data
•Refer to air photos
•Visually selecting
training sites on the
original Landsat TM
images using
human intelligence
Cotton Field
3:37 AM
3. Create Training Areas
1. Create training areas for each category.
2. In ERDAS Imagine, we do this by define
Area of Interest (AOI)
3:37 AM
4. Create Signatures.
A set of statistics that defines the multispectral characteristics of a target
phenomenon or training site.
3:37 AM
5. Choose Best Supervised
Algorithm.
1. Minimum Distance
2. Parallelepiped
3. Maximum Likelihood
3:37 AM
•
with Null class
•
without Null class
Minimum Distance Classifier
• Every pixel is assigned to the
a category based on its distance
from cluster means.
• Standard Deviation is not
taken into account.
• Disadvantages: generally
produces poorer classification
results than maximum
likelihood classifiers.
• Advantages: Useful when a
quick examination of a
classification result is required.
255
191
127
64
0
3:37 AM
64
127
191
255
Parallelepiped
• Pixels inside of the rectangle
(defined in standard deviations),
are assigned the value of that class
signature.
• Pixels outside of the rectangle
(defined by standard deviations)
are assigned a value of zero
(NULL).
255
191
127
•Disadvantages: poor accuracy,
and potential for a large number of
pixels classified as NULL.
• Advantages: A speedy algorithm
useful for quick results.
64
0
3:37 AM
64
127
191
255
Maximum Likelihood With Null Class
• Pixels inside of a stated threshold
(Standard Deviation Ellipsoid) are
assigned the value of that class
signature.
255
• Pixels outside of a stated
threshold (Standard Deviation
Ellipsoid) are assigned a value of
zero (NULL).
191
•Disadvantages: Much slower than
the minimum distance or
parallelepiped classification
algorithms. The potential for a large
number of NULL.
• Advantages: more “accurate”
results (depending on the quality of
ground truth, and whether or not
the class has a normal distribution).
3:37 AM
127
64
0
64
127
191
255
Maximum Likelihood Without Null Class
• Pixels inside of a stated
threshold (Standard Deviation
Ellipsoid) are assigned the value
of that class signature.
• Pixels outside of stated
threshold (Standard Deviation
Ellipsoid) are classified by
minimum distance rules.
255
191
127
• Disadvantages: Slow Algorithm
• Advantages: High accuracy with
no tied or null pixels.
64
0
3:37 AM
64
127
191
255
Unsupervised Classification
• Unsupervised: A method in which the computer separates the pixels in an
image into classes (clusters) with no direction from the analyst. After the
computer has completed the classifying operation the analyst determines the
land-cover type for each class based on image interpretation, ground-truth
information, maps, field reports.etc. , and assigns each class to a specified
category (aggregation).
Data
Raw Imagery (6 bands)
256 Grey Level Values Each
3:37 AM
Information
Classification
Classified Image
80 Classes (Clusters)
Aggregation
Extracted Information
11 Categories
Steps in Unsupervised Classification
1. Define Classification Scheme
2. Configure and Run Classifier
3. Aggregate Classification
4. Label Classes
5. Check Accuracy
3:37 AM
Unsupervised Classification
• Unsupervised classification (commonly referred to as
clustering) is an effective method of partitioning remote
sensor image data in multispectral feature space and
extracting land-cover information.
• Compared to supervised classification, unsupervised
classification normally requires only a minimal amount
of initial input from the analyst. This is because
clustering does not normally require training data.
3:37 AM
Unsupervised
Classification
• Hundreds of clustering algorithms have been
developed.
• ERDAS Imagine uses clustering algorithm ISODATA
(Iterative Self-Organizing Data Analysis)
3:37 AM
Classification
Based on
ISODATA
Clustering