Remote Sensing ' Lecture 9

1
Lecture 9 Content

• Image processing

2

• For every sampled position each sensor detects
  the reflected energy in its particular waveband
  which is converted to a digital number (DN)
• The digital data undergo processing at the ground
  station, to various levels of radiometric and
  geometric correction
• Plotters can convert these corrected DNs to grey
  levels on film (or digitally) to give the
  familiar LANDSAT and SPOT images
• The following slides identifies some common image
  processing processes

3

• Radiometric corrections
• Applied in the first instance to compensate for
  any variation in response or malfunctioning of
  the detectors
• For example in SPOT, it is important that all
  the detectors (3000 or 6000) of the pushbroom
  arrays give equivalent incident energy
• Also useful to apply corrections for haze

4

• Geometric corrections
• Used to correct for geometric distortions of the
  image
• Some distortions are systematic in nature, due
  for example to the forward motion of the
  satellite or to regular variations in scanning
  mirror velocity
• Most noticeable distortion is the west to east
  earth rotation while the satellite is moving from
  North to south
• Corrected by offsetting each scan line or set of
  scan lines slightly, which gives the typical
  parallelogram shape of LANDSAT or SPOT image

5

• Some of the distortions are non-systematic and
  are due to variations in altitude, and velocity
  of the satellite
• Instruments on the satellite and on the ground
  can measure these variations so that the
  necessary corrections can be applied
• These corrections create a new array of picture
  elements (pixels) by resampling the original array

6

• Noise removal
• Unwanted disturbance in image data that is due to
  the limitations in the sensing, signal
  digitizing, or data recording process
• Possible sources are periodic drift, detector
  malfunction, electronic interference, etc.
• Can degrade or mask the true radiometric
  information content
• Objective is to restore as close to the original
  scene
• Can be systematic, random or both

7

• Random noise problems are nonsystematic
  variations in gray levels from pixel to pixel
  called bit errors
• Moving neighborhood windows (3x3) (5x5) are used
• Types of noise
• Striping
• Occurs in multispectral scanners where there are
  variations of the individual detectors ? there
  are several de-striping procedures
• Live drop
• A number of adjacent pixels along a line may not
  contain genuine digital numbers (DN) ? replace DN
  with the average values from the line above and
  below

8

• Contrast Stretching
• The image DNs when first decomposed only occupy
  part of the total possible range of 0 to 255
• Contrast stretching stretches the DN range to fit
  the range of 0-255
• Stretching makes the image easier to interpret

9

10

• Density slicing
• A threshold is set where all values below the
  selected threshold are given the value 0 and all
  those above are given number 1
• This process can be carried further, using
  several thresholds to create groups of values,
  each with a different number. This process of
  reassigning DN values is called density slicing
  or recoding

11

• Spatial filtering
• Also called convolution passes
• Involves a moving matrix of usually 9, 16, or 25
  pixels over the image, so that every image pixel
  in turn becomes the central pixel of the matrix,
  except of course for the edge pixels
• For each position of the matrix, various simple
  mathematical calculations can be performed to
  arrive at a new value for the central pixel
• These new values can be used to create a new
  pixel array and therefore a new image

12

• Popular filters are 3x3 format
• Low pass or smoothing filter
• Calculates the pixel array mean and which
  enhances low spatial frequencies (large features)
  and surpresses high spatial frequencies (lines,
  noise, etc)
• High pass filter
• Subtracts the low pass values from the original
  central pixel values and sharpens the image (also
  called edge enhancement filter)

13

• Textual filter of various kinds
• For example finding the standard deviation of the
  pixel array and applying it to the central pixel
  position, which results in low values for values
  for smooth areas and high values for rough
  textured areas
• Median filter
• Removes noise in the way the low pass filter
  does, but does not blur edges
• Laplacian filter
• identifies edges only

14

• Image classification
• To automatically categorize all pixels in an
  image into classes or themes such as land cover
  or themes
• Types of classifications
• Supervised classification
• The image analyst is in charge of the pixel
  categorization process by specifying (to the
  computer program), numerical descriptors of
  various land cover types present in the scene
• Requires fieldwork and use of other reference
  level information
• More control over the image classification

15

• Unsupervised classification
• Applied in two separate steps
• Groupings or clusters
• -- image data is classified by
  aggregating
• them into the natural spectral
  groupings in
• the scene by using the computer
• Ground reference data
• -- analyst determines the land cover
  identified
• by the clusters or groups by using
  ground
• reference data

16
Unsupervised classification
17
(No Transcript)
18

• The End