ERS186: Environmental Remote Sensing

1
ERS186Environmental Remote Sensing

• Lecture 12
• Classification and Time Series Analysis

2
Classification

• The output of classification is a nominal hybrid
  variable
• Classification is one of the most widely used
  analysis techniques in RS (it is easy to collect
  class data relative to many continuous data).
• Good classification often relies on a good
  understanding of the RT state variables present
  and how they affect a class.
• If two classes have identical RT state variables,
  they can not be distinguished using RS data alone
  (this doesnt stop people from trying, though!)

3
Classification

• Three types of classification
• Supervised
• Requires training pixels, pixels where both the
  spectral values and the class is known.
• Unsupervised
• No extraneous data is used classes are
  determined purely on difference in spectral
  values.
• Hybrid
• Use unsupervised and supervised classification
  together

4
Supervised Classification

• Steps
• Decide on classes.
• Choose training pixels which represent these
  classes.
• Use the training data with a classifier algorithm
  to determine the spectral signature for each
  class.
• Using the classifier, label each pixel in an as
  one of the pre-determined classes (or,
  potentially, an other class).

5
Classifier Algorithms

• There are a LOT of classifier algorithms.
• We will be covering some of these more explicitly
  next quarter, but it is worth covering some of
  them now.
• Table look up
• Parallelepiped
• Minimum distance
• Maximum likelihood

6
Classifier Inputs

• Classifiers are not limited to raw spectral DNs
• Commonly used classifier inputs
• Transformed spectral data vegetation indices,
  endmember fractions, depth of absorption features
• Spatial data distance to certain landscape
  features
• Texture data difference between spectral values
  of a pixel and its neighbors
• Temporal data change detection results

7
Table Look Up

• For each class, a table of band DNs are produced
  with their corresponding classes.
• For each image pixel, the image DNs are matched
  against the table to generate the class.
• If the combination of band DNs is not found, the
  class can not be determined.
• Modification the table can be a range of values!
• Benefits conceptually easy and computationally
  fast.
• Drawbacks every potential combination of band
  DNs and their class is known (or the range)
  matching pixels against spectral libraries is a
  type of table look up.

8
Table Look Up
9
Parallelepiped

• The minimum and maximum DNs for each class are
  determined and are used as thresholds for
  classifying the image.
• Benefits simple to train and use,
  computationally fast
• Drawbacks pixels in the gaps between the
  parallelepipes can not be classified pixels in
  the region of overlapping parallelepipes can not
  be classified.

10
Parallelepiped
11
Minimum Distance

• A centroid for each class is determined from
  the data by calculating the mean value by band
  for each class. For each image pixel, the
  distance in n-dimensional distance to each of
  these centroids is calculated, and the closest
  centroid determines the class.
• Benefits mathematically simple and
  computationally efficient
• Drawback insensitive to different degrees of
  variance in spectral response data.

12
Gaussian Maximum Likelihood

• Gaussian Maximum Likelihood uses the variance and
  covariance between training class spectra to
  classify the data.
• It assumes that the spectral responses for a
  given class fit a normal (Gaussian) distribution.

13
Gaussian Maximum LikelihoodvsMaximum Likelihood

• Both classify according to what is most likely
  according to which class probability density
  curve is higher at a particular DN value.
• But Maximum Likelihood assumes nothing about the
  shape of that probability density curve. It can
  be parametric or non-parametric, Gaussian,
  normal, or whatever - it is whatever you provide.
  
• On the other hand, Gaussian Maximum Likelihood
  assumes that the probability density curves of
  the information classes fit normal (Gaussian)
  distributions.

14
Gaussian Maximum Likelihood

• We assign any given point in spectral space (i.e.
  a set of DN values of a pixel) to a specific
  class. The assignment (classification) is made
  to that class whose probability of occurrence at
  that point in spectral space is highest (most
  likely) - or to Other if the probability isnt
  over some threshold.
• Benefits Uses the class covariance matrix
  representing not only the first but also the
  second order statistics that often contain much
  of the information in RS data.
• Drawbacks
• computationally inefficient,
• multimodal or non-normally distributed classes
  must be split into normally distributed
  subclasses.

15
Decision Tree
16
Other Classifiers
Bad use of neural networks
Good use of neural networks

• Neural networks use machine learning (aka
  artificial intelligence) techniques to classify
  imagery
• Foreground/ background analysis

17
Geometric Correction
Before you classify

• Usually its important to navigate pixels so
  that they represent their correct geographic
  position. Important for
• Classification accuracy
• Overlaying ground truth data
• Overlaying other GIS layers of information
• And thus being able to extract information
• We need to know, for a given pixel in one image,
  what the corresponding pixel is in the 2nd image.
  So we may rubber-sheet both images onto a base
  map.
• Images are rarely distributed geocorrected to
  true ground coordinates, so we have to perform
  this correction ourselves.
• In general, the best we can do for the same
  spatial resolution datasets, is geocorrect to
  within 1/2 of a pixel.

18
Radiometric Correction
Before you classify

• For some analysis techniques (including ones that
  use LSU endmembers), we need to have all of the
  images in the same radiometric scale.
• Relative radiometric normalization (RRN) the
  radiometric (DN, radiance, reflectance) scale is
  slaved to a master image scale.
• Results a target of the same material
  composition will have the same spectral values in
  all images.

19
Principle Components Analysis
Before you classify

• PCA takes multidimensional data and reduces it to
  axes of uncorrelated data.
• The first PCA axis (PC1) will describe the most
  variance in the image data, second axis will have
  the next most variance, etc
• PC1 and PC2 have been found to represent
  unchanging land cover, and PC3 and higher tend to
  represent changing land cover.

20
Principle Components Analysis
Before you classify

• Steps
• Geometrically correct images.
• Perform PCA transformation.
• Interpret PCA images (each image will contain a
  single axis).

21
Principle Components Analysis
Before you classify
PCA Axis 1
PCA Axis 2
PCA Axis 3
PCA Axis 4
22
Principle Components Analysis
Before you classify

• Benefits
• Relatively easy to do, can reduce a large dataset
  into a much smaller dataset.
• Drawbacks
• Interpretation can be difficult.

23
Land Cover Classification
A classification example

• Mapping of species, communities or ecosystems is
  a big part of RS!
• Empirical classification approaches are typically
  used, rather than physical models.
• In general, the higher the sensor resolutions,
  the finer the ecosystem detail.

24
Global Land Cover
A classification example

• Land cover type is a combination of the physical,
  climatic, biotic and human-influenced factors
  present at a specific location.
• Why study global land cover? Land cover
  influences a large number of processes related to
  global climate change.
• Mass/energy exchange between land and atmosphere
• Required to derive essential surface parameters
  for global change models.

25
Global Land Cover
A classification example

• Gopal et al. 1999 used AVHRR NDVI maps, and a
  fuzzy neural network classifier to classify
  global ecosystems. Note how broad the classes
  are.

26
Global Land Cover
A classification example

• Requirements
• Global training data
• Global image coverage
• Images must be mosaicked (stitched) together
  and radiometrically calibrated
• Typically require powerful computers and parallel
  processing algorithms
• Potential exam question what sensors would be
  appropriate (or inappropriate) for doing this
  research? Why?

27
Explicit Use of Time

• Descriptive
• Amount of change / time between images
• Time series analysis
• High temporal resolution and extent
• Longitudinal and survival analysis
• Low temporal resolution and extent

28
Time Series Analysis

• Variation in images across different times
• Cyclical variation (aka seasonal variation)
• LAI in deciduous ecosystems
• Temperature through the year, or throughout the
  day
• Trend variation (aka long term change in mean)
• LAI in human impacted ecosystems across years
• Temperature across years
• Other variation random or non-random changes
  across time

29
Time Series Analysis

• Data requirements
• Fine enough temporal resolution and extent to
  characterize trends and cycles.
• Lots of images that are geometrically and
  radiometrically corrected.

30
Cyclical Variation

• All ecosystems show a degree of cycling, related
  to, among other things, evapotranspiration, day
  length and temperature differences throughout the
  year.
• We want to characterize period, amplitude and
  offset of the cycles.

31
Time Series Analysis

• Elvidge et al. 1998
• Used 64 (!) MSS images in time series analysis
• Elvidge fit a sine function to the data to
  determine cycling (yearly) and then determined
  long term trends.
• Note there are powerful algorithms for
  describing time series data, but RSers dont use
  them!

32
Badhwars use of Sigmoidal Curve to classify crops

• G. Badhwar, c. 1980-1983
• Used two parameter sigmoidal growth curve to
  model crop reflectance vs. date/crop development
  stage
• Used three (or more) Landsat scenes to estimate
  values for the two parameters and the crop
  emergence date.
• Classified the scene using the two parameter
  images as if they were two spectral bands.

Senescence begins
flowering
emergence
Grain fill
Canopy reflectance
Date/development stage