Title: Remote%20Sensing%20Unsupervised%20Image%20Classification
1Remote SensingUnsupervised Image Classification
21. Unsupervised Image Classification
- The process requires a minimal amount of initial
input from the analyst - A numeric operation searches for natural
grouping of the spectral properties of pixels - The analyst determines the information class for
each spectral class after the spectral classes
are formed
31. Unsupervised Classification
- Chain method
- ISODATA
- Spectral mixture analysis
- Object-based image analysis
42. Chain Method
- Pass 1 builds clusters and calculates their mean
vectors - Pass 2 assigns pixels to clusters based on the
minimum-distance rule
5(No Transcript)
6Pass 1. Cluster Building
- R, a spectral radius used to determine whether a
new cluster should be formed (e.g., 15
brightness) -
- N, the number of pixels to be evaluated between
each major merging of the clusters (e.g., 2000)
7Pass 1. Cluster Building
- C, a spectral distance used to determine merging
clusters when N is reached (e.g., 30 brightness) - Cmax, the maximum number of spectral clusters
(categories) (e.g., 20) to be identified
8Pass 1
- The operation evaluates pixels sequentially,
combining successive pixels into a cluster if
their spectral distance lt R - A cluster is complete when N is reached
- If the spectral distance between two clusters is
lt C, the two clusters are merged, until no
clusters with distance lt C - The new mean is the weighted average of the two
original clusters
9Pass 2. Assigning Pixels
- Assigns pixels based on the minimum distance
classifier - Manual modification based on knowledge of the
area, co- spectral plots, and interactive display
10(No Transcript)
11(No Transcript)
123. ISODATA Method
- Iterative Self Organizing Data Analysis
Technique
13ISODATA
- Parameters required
- Cmax, the maximum number of spectral clusters
-
- T, maximum of pixels whose classes are allowed
to be unchanged between iterations - M, the max number of times of classifying pixels
and calculating cluster mean vectors
14ISODATA
- Minimum members in a cluster (). For example,
if the lt0.01, the cluster is deleted - Maximum Std Dev, when a std dev gt specified
Max-std-dev and the number of members gt 2Min
members, the cluster is split
15ISODATA
- Split separation when the value is changed from
0.0, it replaces Std Dev to determine the
locations of the new mean vectors plus and minus
this split separation value - Minimum distance between cluster means. Clusters
with a weighted distance lt this value (e.g., 3.0)
are merged
16ISODATA
- It uses a large number of passes
- The initial means are determined based on the
mean and std dev of each band
17(No Transcript)
18- http//www.youtube.com/watch?vikArEGp-dv0
19Iterations
- Assigns each pixel using the minimum distance
classifier - The second to Mth iteration re-calculate the
mean vectors examine Min members() - Max std dev
- split separation
- Min distance between clusters
- The iteration stops when T or M is reached
20Readings
- Jensen 1996. 2nd Edition or 2005 3rd Edition,
Introductory Digital Image Processing. Prentice
Hall.
214. Classification of Mixed Pixels
- Mixed pixels - when a sensors IFOV covers more
than one land cover feature - e.g. tree leaves, grass, and bare soil
- Depends on the spatial resolution of sensors and
the scale of features - Sub-pixel classification
- - spectral mixture analysis
-
22Spectral Mixture Analysis
- Mixed spectral signatures are compared to pure
reference spectra - The pure signature is measured in the lab,
field, or from images - Assuming that the variation in an image is a
mixture of a limited number of features - Estimates approx proportion of each pure feature
in a pixel
23Spectral Mixture Analysis ..
- Linear mixture models - assuming a linear
mixture of pure features - Endmembers - the pure reference signatures
- Weight - the proportion of the area occupied by
an endmember - Output - fraction image for each endmember
showing the fraction occupied by an
endmember in a pixel
24Spectral Mixture Analysis ..
Gap, water, Mangrove, forest Kemal Gokkaya 2008
25- Tole L., 2008. Changes in the built vs. non-built
environment in a rapidly urbanizing region A
case study of the Greater Toronto Area,
Computers, Environment and Urban Systems, 32(5)
355-364.
26Spectral Mixture Analysis ..
- Two basic conditions
- I. The sum of fractions of all endmembers in a
pixel must equal 1 - ?Fi F1 F2 Fn 1
- II. The DN of a pixel is the sum of the DNs of
endmembers weighted by their area fractions - D?? F1 D??1 F2 D??2 Fn D??nE?
27Spectral Mixture Analysis ..
- One D???equation for each band, plus one ?Fi
equation for all bands - Number of endmembers number of bands 1
- One exact solution without the E term
- Number of endmembers lt number of bands 1
- Fs and E can be estimated statistically
- Number of endmembers gt number of bands 1
- No unique solution
28Spectral Mixture Analysis ..
- Advantages/characteristics
- - a realistic representation of features
- - a deterministic, not a statistic, method
- - fuzzy set theory vs. fuzzy classification
- Disadvantages
- - does not account for multiple reflections
295. Object-based Classification
- Also called object-based image analysis (OBIA)
- vs. per pixel classification
- All classifiers so far consider the spectral
info of a single pixel regardless of its neighbors
30(No Transcript)
31Kutztown GEOEYE-1
Sean Ahearn, Hunter college
325. Object-based Classification ..
- A two-step process
- I. segmentation of the image into objects
- II. Classiication of the objects
- Works at multiple scales and uses color, shape,
size, texture, pattern, and context information
to group pixels into objects
335. Object-based Classification ..
- Two sets of characteristics can be used to
classify the objects -
- The characteristics of the object itself
- (spectral, texture, shape, etc.)
- The relationship between objects
- (connectivity, proximity, etc.)
-
345. Object-based Classification ..
- Advantages
- Disadvantages
- E-cognition, Definiens and Trimble
-
35Readings