Remote Sensing #2

Front 1

sensors

Back 1

 Detectors must be able to sense a wide range of values-snow to
dark volcanic basalt-without becoming saturated
 For sensors with high radiometric resolution a decrease in contrast
may be required if the image DN range exceeds that of the display

Front 2

image contrast enhancement

Back 2

 Utilize full range of video display capabilities
 Imagine 8 bit display – 256 gray scale layers
 Contrast enhancement is done for display purposes only
 Does not affect actual pixel or DN values

Front 3

Contrast Enhancement

Back 3

 Selection of a contrast enhancement
algorithm depends on:
 Sensor radiometric resolution
 Nature of the original (raw) histogram
 Elements of the scene of greatest interest
to user

Front 4

Min-max contrast stretch

Back 4

 Makes full use of range of output device
 Quant is the range of brightness values that can be displayed (i.e., 255)
 Works best with data that have gaussian (normal) distribution
 Expands the image DN range to fill the dynamic range of the
display device (e.g., 0-255)
 Sensitive to outliers (single pixels that may be atypical and
outside the normal DN range)

Front 5

Percentage Linear Stretch

Back 5

 Uses specified Min and Max values that lie in a certain % of
pixels from the mean of the histogram
 A standard deviation from the histogram mean is often used –
Standard deviation Stretch

Front 6

Histogram equalization

Back 6

 Assigns an approximately equal number of pixels to each of
user-specified gray-scale classes
 Greatest contrast is applied to most populated range of DNs.
So, least contrast applied to TAILS of histogram.
 Determine number of output classes and approximate cumulative
percent
 Determine frequency of each unique DN value
 Determine cumulative probability for each DN value
 Using cumulative probability of the DN, assign it to output whose
cumulative percent most closely matches the cumulative
probability

Front 7

Gaussian

Back 7

 Transforms histogram to a gaussian (bell-shaped)
distribution
 High and low ends of the distribution tend to be strongly
enhanced
 Intermediate DN values change relatively little

Front 8

Log Stretch

Back 8

Maximize contrast in dark
part of the histogram

Front 9

Inverse Log Stretch

Back 9

Maximize contrast in
brightest part of the
histogram

Front 10

pixel

Back 10

 Abbreviation of “picture element”
 Smallest 2-D unit of an image
 Value
-“Brightness Value” (BV)
- “Digital Number” (DN)
 Location (x,y)
-column# (x)
- row# (y)
 Quantization
-conversion of electrical signal to digital number
 Radiometric resolution of signal
-Typically in range of 8 – 12 bits
- 8-bit data range from 0 to 255
- 12-bit data range of 0 to 1023

Front 11

ASPRS Guide to Land Imaging Systems

Back 11

 Civil land imaging satellites – resolution >= 59 meters
 Optical ≈ 40 in orbit, 51 countries
-Two major resolution groups
~20 high resolution systems ( 0.5 to 1.8 meter)
~ 24 mid resolution systems ( 2.0 to 39 meter)
-Radar ≈ 10 in orbit, 18 Countries
 Coverage capabilities.
-Hi-res swaths - 8 to 28 km
- Mid-res swaths - 70 to 185 km
 4 privately funded systems in orbit
-3 US and 1 Israeli
~ Hi-res military market
~ 5th commercial system, RapidEye of Germany,
= Broad area of applications with a 5 micro-satellite
constellation
 Planned European satellites
-“Dual Purpose” - data will serve both military and civil users.
~ Tasking will be shared has not been revealed

Front 12

Landsat

Back 12

 Multi-spectral Scanner (MSS)
 Thematic Mapper (TM)
 Enhanced Thematic Mapper (ETM+)
 longest running program
 multi-spectral image data from space
 focus on land resource applications
-satellite: 5000 lbs, 14 by 9 feet (lxw)

Front 13

landsat orbit

Back 13

 Sun-synchronous, near polar
 ~ 705 km altitude
 9:42 am equator crossing
 Each orbit ~ 99 minutes
 14 orbits per day
 Repeat coverage – every
16 days

Front 14

Landsat Worldwide Reference System

Back 14

 Location over earth catalogued by WRS path/row
 Each scene covers 185 km (wide) by 170 km (long)

Front 15

Landsat - Sensors

Back 15

 Return Beam Vidicom (RBV): Landsat 1 – 3
 Multispectral Scanner (MSS): Landsat 1 – 5
 Thematic Mapper (TM): Landsat 4 & 5
 Enhanced Thematic Mapper + (ETM+):
Landsat 6 & 7

Front 16

Multispectral Scanner (MSS)

Back 16

 MSS - longest collection of data by Landsat (1-5)
 Optical-mechanical scanning system
 Oscillating mirror scans across scene with width of
185 km
 Light from scene is reflected through focus optics
 Filters separate beam into separate  bands

Front 17

MSS – Spectral/Radiometric Properties

Back 17

 Spectral resolution
 0.5-0.6m (green)
 0.6-0.7m (red)
 0.7-0.8m (near-infrared)
 0.8-1.1m (near-infrared)
 Radiometric resolution
 Landsat 1-3: 6 bit
 Landsat 4-5: 8 bit

Front 18

MSS – Spatial Resolution

Back 18

 IFOV is 79 meters
 Pixel size: 56 x 79 m

Front 19

MSS – Temporal Resolution

Back 19

 Landsat 1-3: 18 days
 Landsat 4-5: 16 days
 Routine operation ceased in 1992

Front 20

Landsat - Thematic Mapper (TM)

Back 20

 Introduced on Landsat 4 (1982)
 Improvement over MSS:
 Spectral – extended spectral region – visible, NIR,
mid-IR and thermal
 7 Bands vs. 5
 Spatial – 30m vs. 79m (120m for thermal)
 Radiometric – 8-bit vs. 6-bit
 Temporal – 16 day (Landsat 1-3, 18 day)
 Landsat TM 4 & 5
 Sensor Characteristics
 Type of detector
 VIS –NIR
 Silicon
 MIR
 Indium antimonide
 TIR
 Mercury-cadium-telluride

Front 21

TM - Spectral Sensitivity

Back 21

1. 0.45-0.52um- Blue- Water penetration, cultural
features, smoke plumes,
atmospheric haze
2. 0.52-0.63um - Green -Measure peak green reflectance, cultural features
3. 0.63-0.69um -Red -Chlorophyll absorption region, plant species
differentiation, reduced atmospheric effects
4. 0.76-0.90m NIR Vegetation type, biomass,
land/water boundary
discrimination
5. 1.55-1.75m MIR Moisture content, snow vs.
cloud discrimination
7. 2.08-2.35m MIR Vegetation moisture content,
rocks & mineral types
6. 10.4-12.5m TIR Thermal mapping & veg.
stress, soil moisture

Front 22

Landsat 7 - Enhanced Thematic Mapper +

Back 22

 ETM+ introduced on Landsat 7 (1999)
 Similar to Landsat TM 4 & 5
 optical bands (1-5 & 7)
 30m spatial resolution (Bands 1-5 & 7)
 temporal resolution (16 day)
 radiometric resolution - 8 bit
 Thermal band 6 improved from 120m to 60m
 New 15m Panchromatic band added (band 8)
 0.52 to 0.9 m
 Scan line corrector failed in May 31, 2003

Front 23

SPOT Satellite System

Back 23

-French and other European countries
-takes 8.74 spot images to cover same area as TM

Front 24

spot sensors

Back 24

 SPOT 1 – 3
- two High Resolution Visible
(HRV)
 SPOT 4 & 5
-two HRVIR sensors
~ Added SWIR band
- Vegetation sensor
 HRV sensor
-panchromatic
- Multi-spectral
 Vegetation sensor
- Multi-spectral

Front 25

spot - sensors

Back 25

 Uses linear arrays (pushbroom)
 Two CCD arrays
-6,000 detectors
- Linearly arranged
 Improvements
- No mirror that scans back and forth
- Allows for Longer dwell time
- Pixel size is uniform across the swath width

Front 26

HRV - Panchromatic

Back 26

 Panchromatic (PAN)
 Spatial resolution: 10m
 Spectral resolution:
0.51 – 0.73 m

Front 27

HRV

Back 27

 Multispectral (XS)
 Spatial resolution: 20m
 Spectral resolution
 0.50-0.59 m (green)
- 0.61-0.68 um (red)
- 0.79-0.89 um (NIR)
- 1.58-1.75 um (SWIR band added to Spot 4)
= Makes it HRVIR

Front 28

Vegetation Sensor

Back 28

 Spatial resolution: 1.15 km
- 2250 km x 2250 km scene
 Spectral resolution:
- Band 0 0.43-0.47 m
- Band 2 0.61-0.68 m
- Band 3 0.78-0.89 m
- SWIR 1.58-1.75 m
 Temporal resolution: 1 day  Vegetation Index
 Whole Earth 1 Day

Front 29

hyperspectral

Back 29

The simultaneous acquisition of images in many relatively narrow, contiguous and/or non-contiguous spectral bands throughout the the UV, visible and IR portions of the EMR spectrum

Front 30

Advantages - Hyperspectral

Back 30

1) Acquisition of data in hundreds of
spectral bands simultaneously
2) Many surface materials have diagnosticabsorption features

Front 31

what does it look like?

Back 31

AVIRIS image cube...lots of layers
Resolution Attributes
• Spectral Resolution (hundreds of bands)
• 10 nm width
• Spatial Resolution
•20m - 20 km flying height
•5m
• Radiometric Resolution (12 bits)

Front 32

Hyperspectral Instruments

Back 32

 Satellite Based
-Hyperion- On Board EO-1
- Launched Nov., 2001
~ Spatial
= 30m
~ Spectral
= 220 bands
= 10 nm wide
~ Temporal
= 16 day

Front 33

Earth Observing System (EOS)

Back 33

 Large array of instruments to monitor the
earth, ocean, and atmosphere
 Two satellites
- Terra and Aqua
~ Many Sensors
= MODIS, ASTER, MISR, CERES, MOPPIT
 Free data
 Monitor and map a wide array of the
environment
 Develop Products that people can use

Front 34

MODIS

Back 34

• Moderate Resolution Imaging
Spectrometer
(MODIS)
• On board NASA’s Terra and Aqua
satellites
• 36 bands- Visible – Longwave Infrared
•12-bit
• Spatial resolution: 250m – 1000m
•Daily coverage of the entire earth

Front 35

terra instruments

Back 35

ASTER - Advanced Spaceborne Thermal Emission and Reflection Radiometer
CERES - Clouds and the Earth’s Radiant Energy System
MISR - Multi-angle Imaging Spectroradiometer
MODIS - Moderate-resolution Imaging Spectroradiometer
MOPITT - Measurement of Pollution in the Troposphere

Front 36

High Spatial Resolution Sensors

Back 36

Acquisition of data at high
spatial resolution
(<10 meters)
1) Acquires data across few spectral bands
• This is changing
2) Trade off between spatial and spectral resolutions is an important concept

Front 37

GENERIC RESOLUTION ATTRIBUTES

Back 37

• Spatial Resolution: (.46 m – 5m)
-trade off = small swath (8 km – 20 km)
• Radiometric Resolution: (up to 12-bits)
• Spectral Resolution: currently = VIS & NIR

Front 38

High Spatial Resolution Sensor

Back 38

•Airborne Data Acquisition and Registration
•Spatial resolution: 0.6m -1m (depending on flying height)
-swath ~ 1 km x 1 km
•Spectral resolution: Blue, Green, Red & NIR
•Radiometric resolution: 8 bit

Front 39

High Spatial Resolut

Back 39

Spaceborne Sensor: Ikonos
• Spectral/Spatial resolutions:
- 4 Multispectral bands (4m)
~Blue, Green, Red, NIR
- 1 panchromatic band (1m)
- Swath 11 km
• Temporal resolution: < 3 days
• Radiometric resolution: 11-bits
• Cost: $29 per square km2
(ETM is FREE)

Front 40

QuickBird

Back 40

 Spatial/Spectral Resolution
- 0.6m, Panchromatic
- 2.4m, multi-spectral (B,G,R, NIR)
- 16.5 km Swath Width
 Radiometric
- 11 bit
 Temporal
- 1-3.5 day Revisit Frequency
~ Pointable

Front 41

WorldView 2 – Digital Globe

Back 41

 Launched October 8
th
2009
 1.8 m Multispectral
 8 Bands
 .46 m panchromatic
 11 bit
 16.4 km swath width
 1.1 day repeat

Front 42

MODIS land cover mpapting

Back 42

purpose: partitioning image data into thematic categorie (land cover type..i.e. vegatation, urban)
-generally used to make thematic maps

Front 43

Image Classification

Back 43

 Process –
1. Identify classes of interest from classification system
2. Acquire appropriate data – imagery, ancillary, ground reference
3. Preprocess data – radiometric correction, geometric
registration, transform bands
4. Select classification logic/algorithm
5. Extract training sites
6. Select bands or transformation of bands to classify
7. Extract final training sites
8. Classify image
9. Assess accuracy of classified map
10. Distribute results

Front 44

Pattern Recognition

Back 44

 Image classification based on spectral pattern recognition
-Basis of most image classification
-Uses DN values in different bands
~Spectral Signature Concept

Front 45

Hard Classification

Back 45

i.e. one class
 Supervised
~Identification and location of land cover types are know a priori
 Unsupervised
~ Land cover types are not known a priori and computer
generates spectral similar cluster

Front 46

Fuzzy Classification

Back 46

 Takes into account the heterogeneous/imprecise nature of
real world
 Membership in a category is associated with probability

Front 47

Supervised Classification Process

Back 47

1. Select Training Sites for Each Information Class
2. Calculate Statistics for Each Training Site
3. Evaluate “Seperability” of training sites and band
combinations
4. Apply training sites in decision rule to entire image
5. Evaluate output

Front 48

Select Training Sites

Back 48

 Areas with the land cover type of
interest

Front 49

Feature Extraction/Training Sets

Back 49

- Delineating area
 Interactive digitizing, “Area of Interest”
 Select seed pixel and region grow
 Use automated clustering
- Summarize pixels within training set
 Distribution
 Mean, standard deviation, variance, min/max
 Variance-covariance matrix

Front 50

Assessment of Training
Sites

Back 50

~ Evaluate “Seperability” of
training sites and band combinations
 Histograms
 Coincident spectral plots
 2D scatter plots
-Graphing spectral response for two bands
- Elliptical plots
~ Quantitative
1. Transformed Divergence -range 0-2000
2. Jeffries-Matusita (JM)-range 0-1414
 Covariance weighted distance matrix between category
means
 Higher divergence equates to greater “statistical distance”
between means
~Interactive preliminary classification
~ Representative subscene classification
~ Refinement tends to iterative, trial and error

Front 51

decision rules

Back 51

 Once training sites have been determined and
evaluated, classification then proceeds.
 Actual classification, that is assignment of a pixel to
a class in the entire image is based on a decision
rule
1. Minimum Distance to Means
2. Parallelpiped
3. Gaussian Maximum Likelihood
~ Bayesian

Front 52

Decision RulesMinimum Distance to Means

Back 52

~ Process
 Generate mean value for each training site in each band
 Distance between a pixel spectral values and each of the
category means is measured
 Pixel is assigned based on shortest distance
 Threshold can be used to ensure pixels are within some distance
of a mean training site
~ Advantages/Disadvantages
 Computationally simple and efficient
 Disadvantages – insensitive to different degrees of variance
 Not widely used if spectral classes are close to one another and
have high variance

Front 53

Decision Rules-Parallelpiped

Back 53

~Pixel is classified as to whether it falls within specified range
 Defined as rectangles in feature space
 Based on entire distribution (min-max) or plus/minus 2 standard deviations
~ Advantages/Disadvantages
 Includes sensitivity to category variance
 Problems when signatures overlap
 Alleviate with stepped boundaries that parallel the
distribution
 Fast and computationally simple

Front 54

Decision RulesGaussian Maximum Likelihood

Back 54

 Evaluates both variance and covariance of category
spectral response
 Assume normal distribution of pixels in training site
which is described by mean vector and covariance
matrix
 Using these statistics, probability that a given pixel is
member of any one class can be computed
 Pixel assigned to most likely class (highest
probability

Front 55

Decision Rules- Bayesian

Back 55

 Extension of maximum likelihood
 Determine anticipated likelihood of occurrence for
each class in scene (a priori probability)
 Weight associated with cost of misclassifying a pixel
 Generally assumption is equal probability of all
classes unless other info is available
 Maximum likelihood methods are computationally
intensive
~Lookup tables
~ Reduce dimensionality
~ stratification

Front 56

Recap-Supervised Classification

Back 56

~Supervised - Training site issues
 Representative
 Variability
 Signature extension
 Normal/multimodal distribution
 Separability
- Graphical assessment
- Quantitative assessment
~ Decision Rule
 Parallelepiped (Non-parametric)
 Minimum Distance
 Maximum Likelihood (parametric)
~ Output is either in a class or “unknown”
 Distance Image
 Is it correct

Front 57

Unsupervised Classification

Back 57

 Similar to Supervised classification except the human
element occurs later in the process
 Unsupervised Classification Steps
1. Clustering
2. Labeling
3. Evaluate output
 Clustering algorithms
- ISODATA
- K-Means
- RGB

Front 58

ISODATA

Back 58

~ISODATA – Iterative Self-Organizing Data Analysis Technique
 First iteration arbitrarily determines means of N clusters
 After each iteration, a new mean for each cluster is determined
based on actual spectral values
 Used as means for defining clusters in next iteration
 Process continues until little change between iterations
~Advantages
 Not geographically biased to top or bottom of file
 Successful at finding clusters inherent in the data
~ Disadvantages
 Time consuming
 Does not account for pixel spatial homogeneity

Front 59

Unsupervised Classification

Back 59

- ISODATA or other clustering algorithm clusters are
labeled as a class
 Evaluate the Signatures
~ Histograms, Feature Space, TD, JM
 Input Signatures in Classification Decision Rule (PP, Minimum
Distance, Maximum Likelihood)
~ Outputs a classified image
~ Thresholding
 Overlay the classified image pixels on the original image
 If areas cannot be labeled a class
~ a mask is created and the non-classified areas are run again in the
clustering algorithm
- Iterative Process
- Outputs are then recoded into the final map

Front 60

Other Classification Techniques

Back 60

- Remotely sensed information is imprecise:
 Boundaries between phenomenon are fuzzy
 Heterogeneity within classes
- Fuzzy classification
 Pixels are assigned “membership grade” vector
 Describes how close the pixel value is to the mean of
training class vectors
~ Proportion of component classes in a pixel can be
estimated

Front 61

Accuracy Assessment

Back 61

 Error matrix
-For selected sample of pixels compare “actual” category with
classified category
~ Columns = reference data
~ Rows = classification generated from remotely sensed data
 Reference data – “ground truth”
- Should not be used for training sites
 Errors
- Commission – Inclusion
- Omission - Exclusion
 Sampling strategy
 Sample Size