Remote Sensing Part 4

Remote Sensing
Part 4
Classification & Vegetation Indices
Classification Introduction
• Humans are classifiers by nature - we’re always
putting things into categories
• To classify things, we use sets of criteria
• Examples:
– Classifying people by age, gender, race, job/career, etc.
– Criteria might include appearance, style of dress, pitch of
voice, build, hair style, language/lexicon, etc.
– Ambiguity comes from:
• 1) our classification system (i.e., what classes we choose)
• 2) our criteria (some criteria don’t differentiate people with
complete accuracy)
• 3) our data (i.e., people who fit multiple categories and people
who fit no categories)
Non-Remote Sensing Classification
Example
• “Sorting incoming Fish on a conveyor
according to species using optical sensing”
Sea bass
Species
Salmon
** The following data are just hypothetical
Methods
– Set up a camera and take some sample images to
extract features
• Length
• Lightness
• Width
• Number and shape of fins
• Position of the mouth, etc…
Scanning the Fish
• Classification #1
– Use the length of the fish as a possible
feature for discrimination
• Fish length alone is a poor feature for classifying fish
type
– Using only length we would be correct 50-60% of the time
– That’s not great because random guessing (i.e., flipping a
coin) would be right ~50% of the time if there are an equal
number of each fish type
• Classification #2
– Use the lightness (i.e., color) of the fish as a
possible feature for discrimination
• Fish lightness alone is a pretty good feature for
classifying fish by type
– Using only lightness we would be correct ~ 80% of
the time
• Classification #3
– Use the width & lightness (i.e., color) of the
fish as possible features for discrimination
• Fish lightness AND fish width do a very good
job of classifying fish by type
– Using lightness AND width we would be correct
~90% of the time
How does this relate to remote
sensing?
• Instead of fish types, we are typically
interested in land cover
– For example: forests, crops, urban areas
• Instead of fish characteristics we have
reflectance in the spectral bands collected
by the sensor
– For example: Landsat TM bands 1-6 instead
of fish length, width, lightness, etc.
Imagery Classification
• Two main types of classification
– Unsupervised
• Classes based on statistics inherent in the
remotely sensed data itself
• Classes do not necessarily correspond to real
world land cover types
– Supervised
• A classification algorithm is “trained” using
ground truth data
• Classes correspond to real world land cover
types determined by the user
Notes
• For ease of display the following examples show just 2 bands:
– one band on the X-axis
– one band on the Y-axis
• In reality computers use all bands when doing classifications
• These types of graphs are often called feature space
• The points displayed on the graphs relate to pixels from an
image
• The term cloud sometimes refers to the amorphous blob(s) of
pixels in the feature space
Unsupervised Classification
• Classes are created based on the locations
of the pixel data in feature space
255
v
Infrared BV’s
0
0
Red BV’s
255
Unsupervised Classification
A Computer Algorithm Finds Clusters
255
v
Infrared BV’s
0
0
Red BV’s
255
Unsupervised Classification
• Attribution phase – performed by human
255
agriculture
Infrared BV’s
forest
Soil
water
0
0
Red BV’s
255
Problems with Unsupervised Classification
The computer may consider
these 2 clusters (forest and
agriculture) as one cluster
The computer may consider
this cluster (soil) to be 2
clusters
255
v
Infrared BV’s
0
0
Red BV’s
255
Supervised Classfication
• We “train” the computer program using ground truth data
• I.e., we tell the computer what our classes (e.g., trees, soil,
agriculture, etc.) “look like”
Deciduous trees
Coniferous trees
Supervised Classification
Sample pixels
255
v
Other pixels
Infrared BV’s
0
0
Red BV’s
255
Supervised Classification
• No attribution phase necessary because
we define the classes before-hand
255
agriculture
Infrared BV’s
forest
Soil
water
0
0
Red BV’s
255
Problems with Supervised Classification
agri
255
v
forest
v
Infrared BV’s
Soil
water
0
0
Red BV’s
255
What’s this?
What is the computer actually doing?
• This classification
generates statistics
for the center, the
size, and the shape
of the sample pixel
clouds
• The computer will
then classify all the
rest of the pixels in
the image using
these statistical
values
Example: Remote Sensing of Clouds
Supervised Classification: Training Samples
•
Users survey (using GPS) areas of “pure” land cover for all possible land cover types in an image
•
OR
•
Users “heads-up” digitize “pure” areas using expert knowledge and/or higher spatial resolution imagery
•
The rest of the image is classified based on the spectral characteristics of the training sites
Classification of Nang Rong imagery
Shown are Landsat MSS,TM,and ETM
Image Classification Results
(a) Nov 1979
Upland Ag
Forest
Rice
Water
(a) Nov 1992
Built-up
(c) Nov 2001
Land Use/cover Change in Nang Rong,
Thailand
1954
1994
Example Classification Results
(Bangkok, Thailand)
Accuracy Assessments
• After classifying an image we want to
know how well the classification worked
• To find out we must conduct an accuracy
assessment
How are accuracy assessments done?
• Basically we need to compare the classification
results with real land cover
• As with training data, the real land cover data
can be field data (best) or samples from higher
spatial resolution imagery (easier)
• What points should we use for the accuracy
assessment?
– Possible options (there are others)
• Random points
• Stratified random points (each class represented with an
equal number of points)
Classification Challenges
• What problem might occur when gathering
points for an accuracy assessment (and to
a lesser extent, training areas)?
• Can we use the same points for the
accuracy assessment that we used to train
the classification?
Ikonos Imagery: Glacier National Park
Classification Results
Accuracy Assessment Table
•
Rows are the reference data, columns are the classified data
•
Values on the diagonal are correctly classified
•
The values in red are the producer’s accuracy for each class
– A.k.a. errors of omission
– E.g., “how many pixels that ARE water (13) are classified AS water (12)”
•
The values in blue are the user’s accuracy for each class
– A.k.a. the errors of commission
– E.g., “how many pixels classified AS water (14) ARE water (12)”
•
Overall accuracy = # of correctly classified pixels / total # of pixels
•
The Kappa statistic is basically the overall accuracy adjusted for how many pixels we
would expect to correctly classify by chance alone
Vegetation Indices
Vegetation Indices
• Normalized Differential Vegetation Index (NDVI)
• Takes advantage of the “red edge” of vegetation
reflectance that occurs between red and near
infrared reflectance (NIR)
• NDVI = (NIR – Red) / (NIR + Red)
• Many more indices with many variants exist (lots
of acronyms like SAVI, etc.)
Normalized Difference Vegetation Index (NDVI)
RNIR  RRe d
NDVI 
RNIR  RRe d
NDVI: [-1.0, 1.0]
Often, the more leaves of vegetation there are present, the bigger
the contrast in reflectance in the red and near-infrared spectra
NDVI most accurately approximates the Fraction of Absorbed
Photosynthetically Active Radiation (FAPAR)
NDVI from AVHRR
Feb 27-Mar 12
Apr 24-May 7
Jun 19-Jul 2
Jul 17-Jul 30
Aug 14-Aug 27
Nov 6- Nov19
NDVI and Precipitation Relationships
A: 12 Apr-2 May 1982
B: 5 to 25 Jul 1982
C: 22 Sep to 17 Oct 1982
D: 10 Dec 1982-9Jan 1983
Expansion and contraction of the Sahara
Monitoring forest fire
Pre-forest fire
Post-forest fire
Burned area identified
from space using NDVI