Transcript lecture6 - UCL Department of Geography

Remote Sensing and Image
Processing: 6
Dr. Mathias (Mat) Disney
UCL Geography
Office: 301, 3rd Floor, Chandler House
Tel: 7670 4290
Email: [email protected]
www.geog.ucl.ac.uk/~mdisney
1
Image processing: information
extraction - spatial filtering and
classification
Purpose
• To extract useful information from EO data
• Already seen image enhancement and
spectral arithmetic
• Today, spatial filtering and classification
2
Spatial filtering
• Spatial information
– Things close together more alike than things further apart (spatial
auto-correlation)
– Many features of interest have spatial structure such as edges,
shapes, patterns (roads, rivers, coastlines, irrigation patterns etc.
etc.)
• Spatial filters divided into two broad categories
– Feature detection e.g. edges
– Image enhancement e.g. smoothing “speckly” data e.g. RADAR
3
Low/high frequency
DN
Gradual change = low frequency
DN
Rapid change = high frequency
4
How do we exploit this?
• Spatial filters highlight or suppress specific features
based on spatial frequency
– Related to texture – rapid changes of DN value = “rough”,
slow changes (or none) = “smooth”
43
49
48
49
51
43
50
65
54
51
12
14
9
9
10
43
49
48
49
51
210
225
199
188
189
Smooth(ish)
Rough(ish)
Darker, horizontal
linear feature
Bright, horizontal
linear feature
5
Convolution (spatial) filtering
• Construct a “kernel” window (3x3, 5x5, 7x7 etc.) to
enhances/remove these spatial feature
• Compute weighted average of pixels in moving window, and
assigning that average value to centre pixel.
• choice of weights determines how filter affects image
6
Convolution (spatial) filtering
• Filter moves over all pixels in input, calculate value of central
pixel each time e.g.
43
49
48
49
51
43
50
65
54
51
1/9
1/9
1/9
12
14
9
9
10
1/9
1/9
1/9
43
49
48
49
51
1/9
1/9
1/9
210
225
199
188
189
Input image
??
??
??
filter
Output image
7
Convolution (spatial) filtering
• For first pixel in output image
– Output DN = 1/9*43 + 1/9*49 + 1/9*48 + 1/9*43 + 1/9*50 + 1/9*65 +
1/9*12 + 1/9*14 + 1/9*9 = 37
– Then move filter one place to right (blue square) and do same again so
output DN = 1/9*(49+48+49+50+65+54+14+9+9) = 38.6
– And again….. DN = 1/9*(48+49+51+65+54+51+9+9+10) = 38.4
• This is mean filter
• Acts to “smooth” or blur image
37
38.6
38.4
Output image
8
Convolution (spatial) filtering
• Mean filter known as low-pass filter i.e. allows low frequency
information to pass through but smooths out higher frequency (rapidly
changing DN values)
– Used to remove high frequency “speckle” from data
• Opposite is high-pass filter
– Used to enhance high frequency information such as lines and point
features while getting rid of low frequency information
High pass
9
Convolution (spatial) filtering
• Can also have directional filters
– Used to enhance edge information in a given direction
– Special case of high-pass filter
Vertical edge
enhancement filter
Horizontal edge
enhancement filter
10
Practical
• Try out various filters of various sizes
• See what effect each has, and construct your own
filters
– High-pass filters used for edge detection
• Often used in machine vision applications (e.g. robotics and/or
industrial applications)
– Directional high-pass filters used to detect edges of
specific orientation
– Low-pass filters used to suppress high freq. information
e.g. to remove “speckle”
11
Example: low-pass filter
•ERS 1 RADAR image, Norfolk, 18/4/97
•Original (left) and low-pass “smoothed” (right)
12
Example: high-pass edge detection
•SPOT image, Norfolk, 18/4/97
•Original (left) and directional high-pass filter (edge detection), right
13
Multispectral image classification
• Very widely used method
of extracting thematic
information
• Use multispectral
information
• Separate different land
cover classes based on
spectral response
• i.e. separability in “feature
space”
14
Feature space
• Use different spectral response of different
materials to separate
• E.g. plot red against NIR DN values….
15
Supervised classification
• Choose examples of homogeneous
regions of given cover types
(training regions)
• Training regions contain
representative DN values - allow us
to separate classes in feature space
(we hope!)
• Go through all pixels in data set
and put each one into the most
appropriate class
• How do we choose which is most
appropriate?
16
Supervised classification
• Figures from Lillesand, Kiefer and Chipman (2004)
17
Supervised classification
• Feature space clusters
• E.g. 2 channels of
information
• Are all clusters separate?
18
Supervised classification
• How do we decide which
cluster a given pixel is in?
• E.g. 1 Minimum distance to
means classification
• Simple and quick BUT what
about points 1 and 2?
19
Supervised classification
• E.g. 2 Parallelepiped
classification
• We calculate mimumum and
maximum of each training
class in each band
(rectangles)
• BUT what about when a
class is large and overlaps
another?
20
Supervised classification
• E.g. 3 paraellepipeds with
stepped decision boundaries
• Gives more flexibility
• Examples are all 2D BUT
we can extend any of these
ideas into any number of
dimensions given more
spectral channels
• With 3 channels squares
become cubes etc….
21
Supervised classification
• E.g. 4 Maximum
likelihood
• Now we use
probability rather than
distance in feature
space
• Calculate probability
of membership of each
class for each pixel
• Results in “probability
contours”
22
Supervised classification
• Now we see pixel 1 is
correctly assigned to corn
class
• Much more sophisticated
BUT is computationally
expensive compared to
distance methods
23
Unsupervised classification
• As would be expected – minimal user intervention
• We tell it how many classes we expect, then some iterative
preocedure determines where clusters are in feature space
• Assigns each pixel in turn to a cluster, then updates cluster
statistics and repeats…..
– advantage is automatic
– disadvantage is we don’t know what classes represent
24
Unsupervised classification
• E.g. ISODATA (Iterative self-organising data analysis)
algorithm
–
–
–
–
Start with (user-defined number) randomly located clusters
Assign each pixel to nearest cluster (mean spectral distance)
Re-calculate cluster means and standard deviations
If distance between two clusters lower than some threshold, merge
them
– If standard deviation in any one dimension greater than some
threshold, split into two clusters
– Delete clusters with small number of pixels
– Re-assign pixels, re-calculate cluster statistics etc. until changes of
clusters smaller than some fixed threshold
25
Example: 2 classes, 2 bands
DN
Ch 1
Initial cluster
means
a
Assign pixel 1 to
cluster a, 2 to b etc.
DN
Ch 1
Cluster means move towards
pixels 1 and 2 respectively
a
Pixel 2
Pixel 2
Pixel 1
Pixel 1
b
b
DN
Ch 2
DN
Ch 1
All pixels
assigned to a or
b - update stats
New positions of
cluster means
DN
Ch 2
DN
Ch 1
SD of cluster a
too large?
Split a into 2,
recalculate.
Repeat….
New positions of
cluster means
DN
Ch 2
DN
Ch 2
26
Classification Accuracy
• How do we tell if classification is any good?
– Classification error matrix (aka confusion matrix or contingency
table)
– Need “truth” data – sample pixels of known classes
• How many pixels of KNOWN class X are incorrectly classified as
anything other than X (errors of omission)?
– Divide correctly classified pixels in each class of truth data by COLUMN
totals (Producer’s Accuracy)
• How many pixels are incorrectly classified as class X when they should
be some other known class (errors of commission)?
– Divide correctly classified pixels in each class by ROW totals (User’s
Accuracy)
27
Classification Accuracy
Errors of
comission for
class U
Errors of
omission for
class U
28
Post processing?
• Ideally, all classes would be homogeneous
• In reality we get wheat pixels misclassified as “peas”
etc. (accuracy < 100%)
• Pass majority filter over classified image
– E.g. 3 x 3 majority filter assigns output pixel value to be
majority of nearest 8 pixels
Majority
filter
29
Assessed practical
• Classify Churn Farm airborne image based on information
given to you
• Try supervised methods first
– Accuracy? Test using “truth” data provided
– Anything << 80% is not much use….
– What do your training data classes look like in feature space?
• Then perhaps unsupervised
– Accuracy?
•
Warning: don’t include the white text in any of your training data – the pixel
values (255 in all channels) willl TOTALLY screw up your classes!
30