classification - Building The Pride

download report

Transcript classification - Building The Pride

Digital Image Processing

A wide variety of tools that we use to make
remote sensing data provide even more
information.
 Tools include rectification, resampling,
smoothing, edge enhancement, stretching
etc.
 The first tool considered is classification…
Classification
Patterns in digital numbers
=
Patterns on the landscape
These tools use statistical analysis of multispectral images to enhance the information
content provided by RS data.
Creating a thematic map
Multispectral image
classification depends on
the the fact that surface
materials have different
spectral reflectance
patterns…. Different
spectral signatures.
Supervised vs. Unsupervised
In ‘supervised’ classification the interpreter
provides information about the classes he expects
(or wants) to find.
 “Training Sites” are selected on the image to
identify the patterns in spectral space of
classes/features that are to be identified
 Unsupervised Classification… patterns inherent in
the spectral data drive the classification process.

Unsupervised classification
(contd.)

Unsupervised classification can often produce
information that is not obvious to visual
inspection.
 Very useful for areas where ‘ground truth’ data is
difficult to obtain
 Purely spectral pattern recognition
 The critical issue in ALL image classification is to
equate ‘spectral class’ to ‘informational class’!!

“…The trick then becomes one of trying
to relate the different clusters to
meaningful ground categories. We do
this by either being adequately familiar
with the major classes expected in the
scene, or, where feasible, by visiting the
scene (ground truthing) and visually
correlating map patterns to their ground
counterparts…”
The process….

Step one: cluster analysis (Identifying
clusters in the data)
 Step two:Classification of pixels into
classes based on cluster centers
Simple X Y example
(if it were
only this
simple in
reality….)
A 3D version of spectral ‘clusters’… can easily be extended to n
dimensional spectral space.
The clustering process

Virtually all programs use the identical
algorithm “ISODATA”
 Iterative Self-Organizing Data Analysis
(ISODATA) Tou and Gonzalez 1974)
 Begins by assigning class centriods in
statistical space… (random assignment or
some variation)
Cluster analysis….

The input parameters always requested by
ISODATA include
 The initial number of classes
 A class separation distance (a lumping
threshold)
 And the number of iterations (or statistical
threshold) that will define the end of the
process
The first stage….cluster
analysis
At the end of the first stage… nothing exists but a
set of spectral coordinates in n dimensional space
(where n is the number of spectral bands used in
the classification)
 Clusters have been defined based on the number
of cluster centers you start with and the ‘lumping
threshold’ which defines the distance between
centers in spectral space


ERDAS reports this as a .sig file
The ‘right’ number of classes?
How does one select the ‘correct’ or
‘natural’ or ‘right’ number of classes?
 The goal is INFORMATION CLASSES…
not spectral classes…
 “Expert Assessment and Visual
Comparison” (it just looks better!)
 Statistical tools?

Stage 2… putting all pixels
into classes

There are three primary methods for
assigning image pixels to classes
– Minimum Distance to Means (mindist)
– Parallellpiped
– Maximum likelihood classification (maxlik)
The simplest classification…
Minimum Distance to Means
(MINDIST)
Pixels are assigned to a
class based only on the
minimum Euclidian
distance to the closest
cluster center….
Quick and easy but
doesn’t consider
variability in the data
(the density of the
cluster)
Parallepiped classification defines rectangular decision boundaries
around classes…. The size of the rectangular decision boundary is
defined by the variability in the spectral data….
Decision
boundaries are
defined by
variability of
the cluster in
each
dimension
Misidentified pixels…. A common problem
Maximum Likelihood
Classification

This classification is very common
 Used the variance and covariance of the
data to define a ‘probability density
function’ or probability surface….
 (this assumes a ‘normal’ distribution of the
data)
A probability density surface for the sample data set….. Based on the
variability in the cluster, how likely is the inclusion of a given pixel
Probability
contours for
classes in 2
dimensional
space… these
statistical clouds
extend in n
dimensions….
Supervised Classification
Creating statistical clusters based on ‘a
priori’ information
 The interpreter knows what he wants to
find… and creates ‘signature files’ (cluster
centers) from ‘training sites’ on the
image….

Choosing training sites…

Every class has to be fully identified

Training sites should be chosen from all
across the image
 Training sites should avoid edges where
mixed pixels can add uncertainty to the
classified image*
* A tool to accurately classify “mixed pixels” or
highly heterogeneous areas is to choose training
sites within the mixed area… the spectral
signature for this class can be worked with
independently.

Training sites should include 10 to 100
times as many pixels as the total number of
bands being used in the classification… e.g.
for 7 TM bands training sites for each class
ought to contain at least 70 – 700 pixels.

In agricultural applications not uncommon
to have 100+ training sites / class

Polygons vs. “seeds”…. Rather than
delineate the entire polygon, software can
be used to ‘grow’ a training site…