4.3 Digital Image Processing Common image processing image analysis functions: A. Preprocessing B. Image Enhancement C. Image Transformation D. Image Classification and Analysis
4.3 Digital Image Processing
Common image processing image analysis functions:
A. Preprocessing
B. Image Enhancement
C. Image Transformation
D. Image Classification and Analysis
C. Image Transformations
• Manipulation of multiple bands of data
• Generates a ‘new’ image
1. 3 band combinations
2. Spectral ratioing (arithmetic
operations)
Vegetation indices
NDVI
1. 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 ImageryBand 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 7,4,2In a SWIR composite, sensor band 7 is selected from the short-waveinfrared 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.
Blue Green Red Near IR
Water
Blue Green Red Near IR
Agriculture
Urban
Blue Green Red Near IR
2. Spectral ratioing - Using vegetation indices such as NDVI to
study vegetation
• Chlorophyll - Amount of chlorophyll in leaves affects
the spectral signature in the visible.
• Cells known as ‘spongy mesophyll’ are responsible
for reflection of NIR.
– Reflection occurs where the walls of these cells
meet with air spaces inside the leaf.
• Chlorophyll in healthy
vegetation absorbs most of
visible red and blue for
photosynthesis.
• Amount of near infrared
energy reflected is a
function of
– internal structure
– amount of moisture
Vegetation in imagery
Multispectral imagery valuable for study of vegetation.
◦Distinct appearance in certain spectral bands
◦Distinguishes it from other objects in landscape.
Spectral signature varies with species and envir. factors
◦ ID plants in various stages of life cycle or states of health.
Large areas can be studied quickly.
◦Esp. useful in remote areas (tropical rainforest)
◦Possible to obtain accurate quantitative information from
imagery, together with field data.
Vegetation in imagery
Examples;
• Est. # of acres of forest harvested for timber.
• Predict regional or global yields of crops (wheat,
soybeans)
• Est. quantity of phytoplankton in oceans.
• Healthy vegetation - high reflectance in NIR & low reflectance in red.
Landsat Thematic Mapper Imagery
Band Wavelength
1 0.45 to 0.52 Blue
2 0.52 to 0.60 Green
3 0.63 to 0.69 Red
4 0.76 to 0.90 Near IR
5 1.55 to 1.75 Short Wave IR
6 10.40 to 12.50 Thermal IR.
7 2.08 to 2.35 Short Wave IR
• Sometimes air spaces can be filled with water,
thus a plant's state of hydration can
significantly affect the reflectance in NIR.
– Different species have different leaf cell structures,
which affects reflectance of NIR.
• Related factors – leaf size and orientation also
affect reflectance of NIR.
– For example, broad, thin leaves of deciduous plants
are more reflective than needles of coniferous trees.
• Most NIR that is not reflected by leaves is transmitted.
•provides info to analyst
• In a dense forest canopy, leaves underneath often
reflect the energy transmitted by the top layer of leaves.
• So, sections of a forest with a dense canopy will exhibit
higher DN values in the near infrared band than sections
with sparse canopy.
• Differences among plant species;
– amounts of chlorophyll
– different leaf structures, shapes or orientation
– causes species to absorb, reflect, or transmit differently.
• Veg. may have different spectral signature when it is;
– Emergent
– Mature
– Undergoing normal seasonal changes
– Dormant
• Healthy veg. contains more chl. than stressed or diseased.
• Variations in spectral sigs. can be used to study vegetation through image interpretation.
• When leaves lose their chlorophyll in
autumn their spectral characteristics
change.
• Deciduous more reflective in NIR
than conifers.
false-color composite - brightest red near river,
indicating most vigorous vegetation, may be deciduous
trees, shrubs, and grass.
darker red regions surrounding are coniferous forest.
• Vegetative index - calculated (or derived) from
remotely-sensed data to quantify vegetative cover on
Earth's surface.
• Normalized Difference Vegetative Index (NDVI) most
widely used.
Ratio between measured reflectivity in red and near
infrared.
Gives info on absorption of chlorophyll in leafy
green vegetation and density of green vegetation on
the surface.
Also, contrast between vegetation and soil is at a
maximum.
Normalized Difference Vegetation Index Normalized Difference Vegetation Index
(NDVI)(NDVI) has been in use for many years has been in use for many years
to measure and monitor plant growth to measure and monitor plant growth
(vigor), vegetation cover, and biomass (vigor), vegetation cover, and biomass
production from multispectral satellite production from multispectral satellite
data. data.
NDVI is calculated from the visible
and near-infrared light reflected by
vegetation.
Healthy vegetation (left) absorbs
most of the visible light that hits it,
and reflects a large portion of the
near-infrared light.
Unhealthy or sparse vegetation
(right) reflects more visible light
and less near-infrared light.
The numbers on the figure above
are representative of actual values,
but real vegetation is much more
varied.
• NDVI - ratio of red and near infrared (NIR) spectral bands :
– NDVI = (NIR - red) / (NIR + red)
– Resulting index value is sensitive to presence of vegetation on land
surfaces and used to address vegetation type, amount, and
condition.
• Advanced Very High Resolution Radiometer (AVHRR).
– used to generate NDVI images of large portions of Earth on regular
basis to provide global images that portray seasonal and annual
changes to vegetative cover.
• Thematic Mapper (TM and Enhanced Thematic Mapper Plus (ETM+)
bands 3 and 4 also provides Red and NIR measurements:
– NDVI = (Band 4 - Band 3) / (Band 4 + Band 3)
• Primary differences between AVHRR and Landsat
NDVI is resolution.
– AVHRR resolution is 1km and NDVI is 8
km
– Landsat NDVI resolution is 30 m
• AVHRR data - frequent global NDVI products
• Landsat 7 ETM+ data - greater detail covering less
area.
NDVI equation produces values in the range of
-1.0 to 1.0, where vegetated areas will
typically have values greater than zero and
negative values indicate non-vegetated
surface features such as water, barren, ice,
snow, or clouds.
Erdas: Create NDVI IndexNDVI -1.0 to 1.0
Black values = -0.30Whites values = 0.44
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
DN Values
Classes
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.
• Broad Applications:– monitoring deforestation
– impacts on water quality
– document housing density
– urban sprawl
– wildlife habitat and corridors
LULC Maps
• 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