Transcript Spectral Classification Geo410

Thursday 18 February 2010
Lecture 14: Classification
Reading:
Ch. 7.13 – 7.19
Last lecture: Spectral Mixture Analysis
Classification vs.
Spectral Mixture Analysis
In SMA, image pixels were regarded as being mixed from various proportions of common
materials. The goal was to find what those materials were in an image, and what their
proportions were pixel by pixel.
In classification, the image pixels are regarded as grouping into thematically and spectrally
distinct clusters (in DN space). Each pixel is tested to see what group it most closely
resembles. The goal is to produce a map of the spatial distribution of each theme or unit.
Water – group 1
Desert –
group 3
Forests – group 2
AVHRR Images with pixels similar to vegetation
flagged according to distance
at different tolerances D
D=2
D=48
D=60
Multi-unit veg map
What is spectral similarity?
A
dA
Y
Spectral distance: d
Spectral angle: f
x
dB
B
fA
fB
X
Spectral contrast between similar objects is small
Manual Classification
1) association by spectral similarity
of pixels into units
2) naming those units, generally using
independent information
- reference spectra
- field determinations
- photo-interpretation
Seattle
Basic steps in image classification:
1) Data reconnaissance and self-organization
2) Application of the classification algorithm
3) Validation
Reconnaissance and data organization
Reconnaissance
What is in the scene?
What is in the image?
What bands are available?
What questions are you asking of the image?
Can they be answered with image data?
Are the data sufficient to distinguish what’s in the
scene?
Organization of data
How many data clusters in n-space can be recognized?
What is the nature of the cluster borders?
Do the clusters correspond to desired map units?
Classification algorithms
Unsupervised
Supervised
Form Images
Of Data
Separate Data
Into Groups
With Clustering
No
Classify Data
Into Groups
Choose Training
Pixels For
Each Category
Assign Name
To Each Group
Calculate
Statistical
Descriptors
Satisfactory
?
Yes
No
Satisfactory
?
Yes
Classify Data
Into Categories
Defined
Unsupervised Classification:
K-Means algorithm
Pick number of themes;
set distance tolerance
° 1st pixel defines 1st theme
° is 2nd pixel within tolerance?
- YES: redefine theme
- NO: define 2nd theme
° Interrogate 3rd pixel…
° Iterate, using “found”
themes as the new seed
How do you estimate the
number of themes?
- can be greater than
number of bands
Supervised Classification: What are some algorithms?
•Parallelipiped
•Minimum Distance
•Maximum Likelihood
+
x
•Decision-Tree
“Hard” vs. “soft” classification
Hard: winner take all
Soft: “answer” expressed as probability x belongs to A, B
“Fuzzy” classification is very similar to spectral unmixing
Parallelepiped Classifier
Assigns a DN range in
each band for each class
(parallelepiped)
Advantages: simple
Disadvantages: low
accuracy - especially when
the distribution in feature
space has covariance or
dependency with oblique
axes.
Minimum-Distance Classifier
Uses only the mean of each
class. The unknown pixel is
classified using its distance to
each of the class means. The
shortest distance wins.
Decision
boundaries
Maximum Likelihood
The most commonly used classifier used. A pixel is assigned to the class based
on statistical probability.
Based on statistics
(mean; covariance)
A (Bayesian)
probability function is
calculated from the
inputs for classes
established from
training sites.
Each pixel is then
judged as to the class
to which it most
probably belong.
Maximum Likelihood
For each DN ntuple in the image,
1) calculate the distance to each cluster mean
2) scale by the number of standard deviations
in the direction of the ntuple from the mean
3) construct rule images, pixel by pixel for each
cluster, in which the number of std dev’s is
recorded
4) threshold the rule images (null pixels too far
from a cluster)
5) pick best match (least number of std dev’s and
record it in the appropriate pixel of the output
,
image or map
Decision-Tree Classifier
Hierarchical classifier
compares the data sequentially
with carefully selected features.
Features are determined from
the spectral distributions or
separability of the classes.
There is no general procedure.
Each decision tree or set of
rules is custom-designed.
A decision tree that provides
only two outcomes at each
stage is called a “binary
decision tree” (BDT) classifier.
Pre-processing - dimension transformation
One goal: reduce impact of topography on outcome
Line of constant ratio
y
B
x/y
ratioing
NDVI
A
x
y/z
f
Spectral angle
B
A
Validation
Photointerpretation
Look at the original data:
does your map make sense to you?
Confusion matrices
Well-named. Also known as contingency
tables or error matrices
Here’s how they work…
Classified data
Training areas
A B C D E F
A
480
0
B
0
52
C
D
0
0
0
16
E
F
0
0
0
0
480
68
Col sums
5
0
0
20
0
0
All non diagonal elements
are errors
Row sums give
“commission”
errors
Row sums
0
485
0
72
Column sums give
“omission”
errors
Overall accuracy is the
diagonal sum over the
grand total
This is the assessment
only for the training areas
What do you do for the
rest of the data?
1992
Grand sum
p 586, LKC 6th
A basic problem with classification
We tend to want to classify by land use, and from the remote sensing
perspective this may lead to ambiguity
Mystery
pixel X
is found to be
spectrally similar to:
Theme
A=
grass
meadow
= bear
habitat
golf course
≠ bear
habitat
cemetary
≠ bear
habitat
I want to find bears. Bears like meadows. I train on a meadow
(Theme A) and classify an image to see where the bears are. Pixel
X is classified as similar to A. Will I find bears there?
What’s actually on the
ground –
all three look similar (A)
because they are grassy
Maybe not. 1) they might be somewhere else even they like the
meadow. 2) What a meadow is from the RS perspective is a high
fraction of GV. Other things share this equivalence. Therefore, X
may indeed belong to A spectrally, but not according to use.
Thought exercise: what
would you need to do in
order to classify by land
use?
Next class:
Radar