Transcript 8DIPS162015

4.3 Digital Image Processing
Common image processing image analysis functions:
A. Preprocessing
B. Image Enhancement
C. Image Transformation
D. Image Classification and Analysis
3 band combinations
• Significant advantage of multi-spectral imagery is ability
to detect important differences between surface
materials by combining spectral bands.
• Band combinations are created by combining bands of
spectral data to enhance the particular land cover of
interest.
Landsat Thematic Mapper Imagery
Band
Wavelength
1
0.45 to 0.52
Blue
Useful for distinguishing soil from vegetation.
2
0.52 to 0.60
Green
Useful for determining plant vigor.
3
0.63 to 0.69
Red
Matches chlorophyll absorption-used for
discriminating vegetation types.
4
0.76 to 0.90
Near IR
Useful for determining biomass content.
5
1.55 to 1.75
Short Wave IR
Indicates moisture content of soil and veg.
6
10.40 to 12.50 Thermal IR.
Geological mapping, soil moisture, Thermal
pollution monitoring and ocean current studies.
7
2.08 to 2.35
Short Wave IR
Ratios of 5 & 7 are used to map mineral deposits.
Near Infra Red Composite
• Blue visible band is not used and the bands are shifted;
• Visible green sensor band to the blue color gun
• Visible red sensor band to the green color gun
• NIR band to the red color gun.
• Results in the familiar NIR composite with vegetation
portrayed in red.
Bands 4, 3, 2
Near Infrared Composite (4,3,2)
• Vegetation in NIR band is highly reflective
• Shows veg in various shades of red
• Water appears dark due to absorption
Popular band combination for vegetation studies, monitoring
drainage and soil patterns and various stages of crop growth.
• Vegetation - shades of red
– Conifers darker red than hardwoods
– lighter reds = grasslands or sparsely vegetated
• Urban - cyan blue, light blue
• Soils - dark to light browns.
• Ice, snow and clouds - white or light cyan.
Bands 3,2,1
True Color composite
Visible bands are selected and assigned to their corresponding
color guns to obtain an image that approximates true color.
Tends to appear flat and have low contrast due to scattering of
the EM radiation in the blue visible region.
3, 2, 1
• Ground features appear in colors similar to their appearance
– healthy veg = green
– cleared fields = light
– unhealthy veg = brown & yellow
– roads = gray
– shorelines = white
• Water penetration - sediment and bathymetric info
• Used for urban studies.
• Cleared and sparsely vegetated areas are not as easily detected
• Clouds and snow appear white and are difficult to distinguish.
Bands 7,4,2
In a SWIR
composite,
sensor band 7 is
selected from the
short-wave
infrared region.
• Shortwave Infrared Composite (7,4,3 or 7,4,2)
• SWIR composite image contains at least one
shortwave infrared (SWIR) band.
• Reflectance in SWIR region due primarily to moisture
• SWIR bands are especially suited for camouflage
detection, change detection, disturbed soils, soil type,
and vegetation stress.
• Provides a "natural-like" view, penetrates atmospheric particles and smoke.
– Healthy veg = bright green
– Barren soil = Pink
– Sparse veg = oranges and browns
– Dry veg = orange
– Water = blue
– Sands, soils and minerals - multitude of colors.
– Fires = red - used in fire management
– Urban areas = magenta
– Grasslands - light green.
– Conifers being darker green than deciduous
• Provides striking imagery for desert regions
• Useful for geological, agricultural and wetland studies
Use the spectral profile tool (Raster  Profile Tool) to examine the different
spectral properties of a. water, b. vegetation and c. urban areas. Choose several
pixels from each of the 3 categories and plot them.
Water
Blue
Green
Red
Near IR
Agriculture
Blue
Green
Red
Near IR
Urban
Blue
Green
Red
Near IR
D. Image Classification and Analysis
• Process of categorizing all pixels in an image into
land cover classes.
• Multispectral imagery is used.
• Spectral Signatures for each pixel is the numerical
basis for the algorithm.
Continuous data
• Raster data that are quantitative (measuring a
characteristic) and have related, continuous values,
such as remotely sensed images (e.g., Landsat,
SPOT).
Thematic data
• Raster data that are qualitative and categorical.
• Classes of related information, such as land cover,
soil type, slope.
Image data classification
• Primary component of image interpretation
– using computer software to spectrally categorize data
– computer id’s clusters of spectrally similar pixels
– Analyst's knowledge
• how to classify the image data
• assign appropriate descriptions to the categories
• Individual pixels in a continuous image are assigned to
classes.
• Result is a thematic image where each class represents a
feature type in the real world.
•
Create thematic image from multi-spectral continuous
image
Classes
DN Values
unsupervised - analyst may have little knowledge
of what data represents.
supervised - a priori knowledge required.
Each pixel in image contains
information about the surface
materials that reflected light
from that pixel to the sensor.
Each pixel contains a value
which can range from 0 to 255,
for each band in image.
Vegetation Features that are
indistinguishable in
visible region of EMS
can be separated in
near IR.
VIS
NIR
Supervised and Unsupervised
Classification
• Two different approaches to classifying an image
• Each has advantages and disadvantages
• Unsupervised classification
• primarily a computer process
• minimal user input
• analyst assigns an identification to each class,
based on knowledge of the image's content
Supervised classification
• user-controlled process
• depends on knowledge and skills of analyst
for accurate results.
• analyst knows beforehand what feature
classes are present and where each is in
one or more locations within scene.
• Used to train computer to find spectrally
similar areas.
• Unsupervised classification - used to
generate a set of classes for entire
image and make a preliminary
interpretation.
• Then supervised classification can be
used to redefine the classes as more
information becomes available.
ISODATA clustering algorithm
• Unsupervised classification of remote sensing data.
• Uses a minimum spectral distance formula to form clusters.
– begins with arbitrary cluster means, or means of an existing signature
set
– each time clustering repeats, means of the clusters are shifted.
– new cluster means are used for the next iteration.
• Algorithm repeats the clustering of the image until either;
– maximum number of iterations
– maximum percentage of unchanged pixels has been reached between
two iterations.
Ground Truthing
• Verifying that feature classes derived from image
data accurately represent real world features.
• Requires collecting ground truth data.
• Derived from a variety of sources.
– onsite visits, aerial photography, maps, written
reports and other sources of measurements
• Ideally, should be collected at the same time as the
remotely sensed data.
• aerial photos for
ground truthing.
• Amount and type of ground truth
required depends on the level of detail
in the classification.
• Ground truthing can be used to select
training sites prior to supervised
classification or to identify key classes
after unsupervised classification.
• Landsat data (resolution of 30 meters)
is appropriate for classifying general
landscape characteristics across large
areas.
Uses of classification
• Creation of land use and land cover (LULC)
maps.
• Land cover - natural and human made
features: forest, grasslands, water and
impervious surfaces.
• Land use - how land is used: protected area,
agricultural, residential, and industrial.
• LULC classification
system - widely
used as a general
framework.
LULC Maps
• Broad Applications:
– monitoring deforestation
– impacts on water quality
– document housing density
– urban sprawl
– wildlife habitat and corridors
• Land cover classification/change detection analysis
for the Columbia R. coastal drainage area
Unsupervised classification
• classes are determined by software based on
spectral distinctions in data
• little knowledge of imaged area is required
• To assign identification to each class requires
some knowledge of the area from personal
experience or from ground truth data.
• primary advantage - distinct spectral classes
are identified.
• Many of these classes might not be initially
apparent to the analyst.
• Spectral classes may be numerous.
Unsupervised
• Primary disadvantage - spectral patterns
identified by computer do not necessarily
correspond to meaningful features of land
cover or land use in the real world.
Supervised Classification
• Classes determined by analyst.
• Use pattern recognition skills and prior
knowledge of the area to help software
determine spectral signatures for each
class.
Supervised classification
• More accurate than unsupervised
classification, provided that the classes are
correctly identified by the analyst.
• Disadvantage - accurately establishing the
classes can be a very time-consuming
process.
TRAINING SITES
• Critical part of supervised classification.
• Includes spectral characteristics for each
land cover type to be classified in an image.
• Software uses them to find similar areas
throughout the image.
• May need to establish several training sites
for each class.
4 training sites to establish agriculture class.
Unsupervised Classification 6 Classes
Save as a PDF