Introduction to Remote Sensing Lecture 11

1
Introduction to Remote SensingLecture 11
2
Remote Sensing

• So far, we have aimed to answer the following
  questions
• Why use the technique (remote sensing)?
• What is the physical basis ?
• How are the data collected ?
• What range of sensors are there?
• How can we enhance data ?
• The question to be answered in the next 2
  lectures is
• How can we produce thematic maps ?

3
(No Transcript)
4
Image Classification
5
Image Classification

• Image Classification uses the spectral
  information represented by the digital numbers in
  one or more spectral bands, and attempts to
  classify each individual pixel based on this
  spectral information
• The objective is to assign all pixels in the
  image to particular classes or themes (e.g.
  water, coniferous forest, deciduous forest, corn,
  wheat, etc.).
• The resulting classified image is comprised of a
  mosaic of pixels, each of which belong to a
  particular theme, and is essentially a thematic
  "map" of the original image.

6
Image Data
Thematic Map
Image Classification
7
Spectral or Information Classes ?

• When talking about classes, we need to
  distinguish between
• Information classes (e.g. land use)
• Spectral classes (e.g. land cover)

8
Information Spectral Classes

• Information classes are those categories of
  interest that the analyst is actually trying to
  identify in the imagery, such as different kinds
  of crops, different forest types or tree species,
  different geologic units or rock types, etc.
• Spectral classes are groups of pixels that are
  uniform (or near-similar) with respect to their
  brightness values in the different spectral
  channels of the data.
• The objective is to match the spectral classes in
  the data to the information classes of interest.

9
Information Spectral Classes

• Rarely is there a simple one-to-one match between
  these two types of classes.
• Rather, unique spectral classes may appear which
  do not necessarily correspond to any information
  class of particular use or interest to the
  analyst.
• Alternatively, a broad information class (e.g.
  forest) may contain a number of spectral
  sub-classes with unique spectral variations.
• It is the analyst's job to decide on the utility
  of the different spectral classes and their
  correspondence to useful information classes.

10
Technicality

• Assignment of spectral classes to information
  classes
• A key process in land-cover mapping is the
  aggradation of spectral classes, and their
  assignment to information classes (especially in
  the case of unsupervised methods)
• For example, accurate classification of the class
  deciduous forest may require several spectral
  sub-classes, such as north-facing forest,
  south-facing forest, shadowed forest, and the
  like.
• When the classification is complete, these
  sub-classes should be assigned a common symbol to
  represent the single informational class

11
(No Transcript)
12
(No Transcript)
13
Case Study Churn Farm
14
Churn Farm

• Remote Sensing data were collected on 4 June 1984
  using the NERC ATM scanner for Churn Farm,
  Berkshire.
• The ATM scanner that was used has 12 optical
  bands.
• Seven of these correspond to Landsat TM
  wavelengths, the other 5 are experimental bands.
• The aeroplane carrying the ATM scanner was flown
  at a low altitude and the image has a spatial
  resolution of 5 meters.

15
Daedalus ATM bands (7,4,2)
16
Technicality

• Selection of Images
• Generally, the success of a land cover
  classification can lie in the astute selection of
  imagery with respect to season and date.
• Therefore the seemingly mundane process of
  searching image archives for suitable data
  assumes vital significance......stemming from the
  need to answer questions such as
• What season will provide the optimum contrasts
  between classes to be mapped?

17
Technical Detail

• Details about the Churn Farm data
• 1. Collected 4 June 1984
• 2. Grass appears to consist of 2 different
  varieties, and 2 of the smaller wheat fields look
  distinct from the rest.
• 3. The peas have only recently been planted, and
  much bare soil will be showing through.
• 4. There is a small amount of cloud in the upper
  left part of the scene.
• 5. The Urban areas consist mainly of farm houses
  and farm yards there is also an electricity
  sub-station. In most of these units, there will
  be several pixels that are pure grass or trees.

18
(No Transcript)
19
Information Classes Spectral Classes ? 1.
Winter wheat 1. Wheat 2. Winter barley 2.
Barley 3. Winter beans 3. Beans 4. Peas
4. Peas 5. Lucerne 5. Lucerne 6. Ley
Grass 6. Grass 7. Gallops 8. Pasture 9.
Other Grass 10 Scrub 7. Scrub 11.
Plantation 8. Trees 12. Trees 13. Buildings
9. Urban 14. Metalled Roads 15. Concrete
Tracks 16. Tracks and paths
20
(No Transcript)
21
Classification Types

• Common classification procedures can be broken
  down into two broad subdivisions based on the
  method used
• Supervised classification and
• Unsupervised classification

22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
Technicality

• Land cover can be mapped from remote sensing
  using a range of classification techniques.
• In principal, the process is straightforward in
  practice, many of the most significant factors
  are concealed among apparently routine
  considerations.

26
Supervised Classification
27
Supervised Classification

• Objective To automatically categorize all pixels
  in an image into information classes.
• requires an analysis of the spectral properties
  of surface features in a multi-band image and
• a systematic sorting, based on mathematical
  decision rules, of the spectral data into
  spectral/textural categories.
• Assumption different surface features manifest
  different combinations of digital values based on
  their spectral reflectance/emittance/backscatter
  properties.

28
Supervised Classification

• Definition of Information Classes
• Training/Calibration Site Selection
• delineate areas of known identify on the digital
  image
• Generation of Statistical Parameters
• define the unique spectral characteristics
  (signatures)
• Classification
• assignment of unknown pixels to the appropriate
  information class
• Accuracy Assessment
• test/validation data for accuracy assessment
• Output Stage

29
Training Stage

• Objective
• To assemble a set of statistics that describe the
  spectral response pattern for each information
  class.
• Involves the delineation of areas of known
  identify on the digital image.
• Requires
• close interaction between the analyst and the
  image data
• reference data

30
The Training Stage

• For an accurate classification,
  training/calibration data must be
• Representative
• to adequately sample the spectral variation for
  the information class
• sample numerous small areas scattered throughout
  the image
• Complete
• a sufficient sample size is required to ensure
  accurate statistical descriptors

31
Oblique Air Photo of Morrow Bay, California
32
TM Band 4 red TM Band 3 green TM Band 2
blue
TM Band 3 red TM Band 2 green TM Band 1
blue
Landsat TM, Morrow Bay, California
33
Training Stage - Technicality

• In reality, therefore, it can be useful if the
  operator is familiar with the location from which
  the remotely sensed data has been acquired.
• This will make the selection of training sites
  relatively straightforward.
• In addition, any in-situ spectral measurements of
  the training areas taken at the time of data
  collection will be taken into account.

34
Training Sites Factors to Consider (1)

• Number of Training/Calibration Areas
• depends on (i) of classes and (ii) diversity
  of classes
• many smaller areas better than a few large areas
• Number of Training/Calibration Pixels
• 10N to 100N pixels where n of spectral bands
  (Lillesand and Kiefer)
• depends on the environment, but at least 100
  pixels per class accumulated from several
  training areas (Campbell)
• gt10N pixels where n no. of spectral bands
  (Jensen)

35
Training Sites for Land-Cover Units
36
Training Sites Factors to Consider (2)

• Size
• large enough to provide accurate estimates of
  each information class
• not too large to result in undesirable variation
• Shape
• not important
• squares, rectangles (right-angled shapes easy to
  work with)

37
Training Sites Factors to Consider (3)

• Location / Placement
• several training areas throughout the image (use
  maps and airphotos if field visit not possible
• relate to recognizable ground features
• keep away from boundaries
• Uniformity
• should be homogeneous (unimodal) rather than
  heterogeneous (bimodal or multimodal)

38
When DNs are plotted as a function of the band
sequence (increasing with wavelength), the result
is a spectral signature or spectral response
curve for that training class. In reality the
spectral signature is for all of the materials
within the training site that interact with the
incoming radiation.
39
Technicality

• Selection of Training Data
• Accurate selection of training data is crucial
  for accurate supervised classification.
• There are many approaches to signature collection
  and analysis, but all rely to a certain degree on
  the experience of the analyst.

40
Final Product