Transcript Classification of Remotely Sensed Data

Classification of Remotely
Sensed Data
General Classification Concepts
Unsupervised Classifications
Learning Objectives
• What is image classification?
• What are the three broad classification
strategies?
• What are the general steps required to classify
images?
• What is a classification scheme, and how is a
good one constructed?
• What is unsupervised classification, and what
are its advantages and disadvantages?
Wyoming land cover modified from
USGS Gap Analysis Program data
•
•
•
Based on Landsat TM data
Used a supervised classification
technique called CART analysis
Legend is aggregated from much
more detailed legend
USGS National Land
Cover Database
(2006)
•
•
•
Based on Landsat
TM data
Used CART
Legend is not
aggregated – not as
detailed as GAP
U.S. Landcover - MODIS
What is Image Classification?
• Process of grouping image pixels into classes
that represent self-similar features or themes
Analogous to any
classification exercise.
How?
• Three general approaches:
– Manual interpretation (e.g., photointerpretation,
“heads-up digitizing”)
– Unsupervised classification of digital data
– Supervised classification of digital data
Sequence of methods similar for all three!
A word about manual classification
• Manual image interpretation was at the core
of remote sensing for much of its history
• Still perfectly appropriate today in some
situations
• Usually requires people trained in
photointerpretation to make decisions about
boundary placement and class labels
• Can use computers for on-screen
interpretation
Detailed view of Wyoming GAP Land Cover Map
General Classification Steps
1.
2.
3.
4.
Field reconnaissance
Development of a classification scheme (legend)
Image enhancement/preprocessing (as needed)
Classification using manual or digital techniques
Incorporation of ancillary data (as needed)
5. Accuracy assessment
6. Iterative refinement
From: http://gisgeography.com/image-classification-techniques-remotesensing/
Field Reconnaissance
• Critical for understanding the distribution of
your theme in the real world
– Understanding of ecology, geology,
geomorphology, etc.
• Helps you choose useful ancillary data
• Useful for interpreting the imagery
• Nice excuse to get out of the office once in a
while
What characteristics of this landscape might be important for making a
landcover map using satellite data?
Developing Classification Schemes
(Legends)
• How many types do you want to map?
• How should you divide up the feature you are
interested in?
• Can be very controversial!
Classification Schemes
(List of types to map)
1)
2)
3)
4)
5)
Must be useful (how will map be used?)
Must be detectable using the data you have
Usually hierarchical
Categories must be mutually exclusive
Require explicit definitions of each class
Classification Scheme -- Example
I.
Vegetated
A.
Forest
1.
Evergreen
a.
b.
c.
2.
B.
II.
Spruce-fir forest
i. Spruce-fir with winterberry
understory
Lodgepole pine forest
etc.
Deciduous
Shrubland
Non-Vegetated
Groups
Generate a classification scheme
for mapping the main UW campus
based on Google Earth imagery
Digital classification
• Creating thematic classes based on groups
of similar digital numbers (DNs)
– Statistical grouping of the data (puts spectrally
similar pixels into the same class)
– Spectral vs. informational classes
– Sometimes combine spectral classes to make
informational classes
Classification
Use many bands at once to create a map of classes
Classification Strategies
• Unsupervised – computer clusters pixels
based only on the similarity of their DNs.
• Supervised – computer uses training data—
examples of target classes—and assigns pixels
to the closest training class using similarity
– Others (neural networks, fuzzy logic etc.)
Max
Unsupervised Classification 2-bands
Band Y
.
... .
. .
. .. .
..
.
.
. ..
..
0
..
. . ..
. ..
.
. .. . .. . .
... .
..
.. . . . .
. . . ..
.
..
.
.. . ..
. .. .
.
.
.
. .
.. .. . . .
.. .
..
.
.
.
.
. .. . .
. .. . . .
.
. ... . .
. .. . . . . . . . . . .
..
Band X
Class
Max
Unsupervised Classification 3-bands
1 pixel
1 Class
Unsupervised Classification
• Decide how many groups (classes) you want
• Choose bands, indices, enhancements, etc.
that highlight differences in your classes
• Choose a grouping algorithm
– Simple clustering, K-means, etc.
• Classify the image (run the algorithm)
• Label (and aggregate) classes and evaluate the
results
Advantages of Unsupervised Classifications
• No extensive prior knowledge of map area
required (but you have to label the classes!)
• Classes are based only on spectral information
(spectral classes), so not as subjective as when
humans make decisions
• Can use results of unsupervised classification
to help guide you in the field, even if
ultimately you use supervised classification.
Disadvantages of Unsupervised
Classifications
• Spectral classes do not always correspond to
informational classes
• Spectral properties of informational classes
change over time so you can’t always use
same class statistics when moving from one
image to another
Grouping Algorithms
• Statistical routines for grouping similar pixels
• Operate in feature space
• Differ in how they:
– Determine what is similar (distance measures)
– Determine the statistical center (centroid) of a
class
Unsupervised classification of Martian terrain.
From Stepinski and Bagaria. 2009. IEEE Geoscience and Remote Sensing Letters.