1 Lectures 11 Analysis and Classification of Digital MSS Data 21 October 2008 Today’s Lecture will...

1

Lectures 11Analysis and Classification of

Digital MSS Data 21 October 2008

Today’s Lecture will be given by Dr. Tatiana Loboda, a Research Scientist in the Department of

Geography

2

Readings

Campbell, Chapters 12,17.9-17.10

3

Challenge in remote sensing – how does one capture the information content that is available in the different channels of the digital image

Fig 1

4snow

lake wateraspen

spruce

5

Snow

0.0

25.0

50.0

75.0

100.0

125.0

150.0

Band 1 Band 2 Band 3 Band 4 Band 5 Band 7

Dig

ital

Nu

mb

er

Lake Water

0.0

25.0

50.0

75.0

100.0

125.0

150.0


Dig

ital

nu

mb

er

Spruce

0.0

25.0

50.0

75.0

100.0

125.0

150.0

Band 1 Band 2 Band 3 Band 4 Band 5 Band 7D

igit

al N

um

ber

Aspen

0.0

25.0

50.0

75.0

100.0

125.0

150.0


Dig

ital

Nu

mb

er

Question: Can we train the computer to use the differences in DN values between the different channels to discriminate the different surface categories?

Fig 3

6

Lecture Topics1. Goals for Image Processing2. Steps for producing an information product from

satellite data3. Definition of image classification4. Image slicing or thresholding

– Thresholding of digital values– Thresholding of transformed values

5. Image categorization or classification– Supervised vs. unsupervised– Multiband Classification

6. Indices generated from multi-band data– Vegetation indices– Burn/fire severity indices

7. Global NDVI data sets8. Analysis of temporal trends in vegetation indices9. Decision tree classifiers

7

Goals for image processing

To identify and quantify features of interest with a specified degree of accuracy which allows for systematic analysis of remotely sensed data

8

Single Characteristi

c

Multiple Categories

Different levels of a single

characteristic

Vegetation categoriesAreas affected by fire

Chlorophyll levelsFig 4

9

Goals for image processing – quantify

1. Create maps with discrete categories -Example - estimate areas with different land cover categories in an urban region

2. Identify and map a specific feature of interest on different dates

– Example - determine areas and rates of deforestation in tropical regions

3. Create a maps that represents different levels of a surface characteristic

– Example - estimate net primary production in oceanic regions

10







11

Steps for producing a satellite-based information product

1. Data exploration (contrast stretching and multi-channel displays)

a. Are you seeing the features within the imagery that were expected?

b. Are their any unusual or unexpected signatures?

c. Do these new signatures represent important information?

12


2. Data correction/normalization**a. Geometric rectificationb. Radiometric normalization –

conversion of DNs to a relative or absolute radiance value

**Topics for GEOG472

13


3. Initial information product generation4. Product evaluation/validation5. Refinement of product generation

approach6. Generation of final product

Focus of Lecture 10 is on different approaches used to generate information products from VIS/RIR multi-channel data – step 3

14







15

Image Classification

The process of automatically dividing all pixels within a multichannel digital remote sensing dataset into

1. Discrete surface-cover categories2. Information themes or quantification

of different levels of a specific surface characteristics

16From: http://www.fes.uwaterloo.ca/crs/geog376.f2001/ImageAnalysis/ImageAnalysis.html#ImageProcessingSteps

Regional land cover map derived from Landsat TM data

Fig 5

17

1: Evergreen Needleleaf Forests; 2: Evergreen Broadleaf Forests; 3: Deciduous Needleleaf Forests; 4: Deciduous Broadleaf Forests; 5: Mixed Forests; 6: Woodlands; 7: Wooded Grasslands/Shrubs; 8: Closed Bushlands or Shrublands; 9: Open Shrublands; 10: Grasses; 11: Croplands; 12: Bare; 13: Mosses and Lichens

http://www.geog.umd.edu/landcover/8km-map.htmlGlobal Land cover map generated from AVHRR dataFig 6

18Ocean chlorophyll concentration derived from SeaWifs data

Fig 7

19

Fig 8

Land cover map generated from Shuttle Imaging Radar data

20

Fig 9

Map of areas impacted by fire generated from MODIS thermal IR data

21

Lecture Topics1. Goals for Image Processing2. Steps in producing an information product from






22

Digital Numbers from Landsat ETM Data

0102030405060708090

100110120

Wat

erBurn

Spruce Soil

Grave

l

Aspen

Snow

Barle

yBan

d 4

(.7

6 to

.90

um

) D

igit

al N

um

ber

Thresholding is carried out on a single channel of data

Fig 10

23

14 15 17 15

16 13 18

16

17

15

15

14

13 15

11 12

Average = 14.75Range = 11 - 18

Most land surfaces have a range of values, not a single value –

Fig 11

24

Lillesand and KieferFigure 7-46

Fig 12

25

Lillesand and KieferFigure 7-11 Fig 13

26


Water

Land

The range in digital values for these two surfaces do not overlap, so you can use a level slice to classify your image into two categories

> 40 = Land< 40 = water

Fig 14

27

Thresholding/Level Slicing Limitations

• When you have a variety of different types of land surfaces, using level slicing with a single channel usually only provides a small number of discrete categories

28







29

Two-step level slicing or thresholding

Step 1 – Estimate a surface characteristic based on the relationship between that characteristic and the different bands of the MSS system

Step 2 – create discrete levels of the characteristic, e.g., level slice the resultant single band channel

30

Example of 2-step level slice

With AVHRR data, greenness can be estimated from the Normalized Difference Vegetation Index (NDVI)

NDVI = (Near IR – Red)(Near IR + Red)

31

This greenness map was created by level slicing NDVI ValuesFig 15

32

Example of 2-step level slice

With Seawifs data, chlorophyll concentration is estimated as a function of several bands

Chlorophyll concentration = f (B1, B2, B3)

33

Chlorophyll concentration based on Seawifs algorithm

Fig 16

34







35

Snow

0.0

25.0

50.0

75.0

100.0

125.0

150.0


Dig

ital

Nu

mb

erLake Water

0.0

25.0

50.0

75.0

100.0

125.0

150.0


Dig

ital

nu

mb

er

Spruce

0.0

25.0

50.0

75.0

100.0

125.0

150.0


Dig

ital

Nu

mb

er

Aspen

0.0

25.0

50.0

75.0

100.0

125.0

150.0


Dig

ital

Nu

mb

er

Fig 17

36

0.0

25.0

50.0

75.0

100.0

125.0

0.0 25.0 50.0 75.0 100.0 125.0

Band 3 (.60 to .69 um) digital number

Ban

d 4

(.7

6 to

.90)

dig

ital

nu

mb

er

snow

water

burn

aspen

spruce

gravel

soil

barley

Fig 18

37


Fig 19

38







39

Supervised versus Unsupervised Classification

• Supervised classification – a procedure where the analyst guides or supervises the classification process by determining numerical descriptors of the land cover types of interest

• Unsupervised classification – the computer is allowed to aggregate groups of pixels into clusters with similar digital values

40

Training Areas and Supervised Classification

1.Specified by the analyst to represent the land cover categories of interest

2. Used to compile a numerical “interpretation key” that describes the spectral attributes of the areas of interest

3. Each pixel in the scene is compared to the training areas, and then assigned to one of the categories

41







42

Multiband Classification Approaches

– Minimum distance classifiers– Parallelepiped classifiers– Maximum likelihood classifiers– Decision tree classifiers – see

lecture topic 9

43

Minimum Distance Classifiers

f

f

f

ff

Step 1 – calculate the average value for each training area in each band

+

cc

c

cc+*

* - Unclassified pixel

44


f

f

f

ff

Step 2 – for each unclassified pixel, calculate the distance to the average for each training areaThe unclassified pixel is place in the group to which has the smallest value+

cc

c

cc+*

45



Fig 20

46

Advantages/Disadvantages of Minimum Distance Classifiers

• Advantages– Simple and computationally efficient

• Disadvantages –– Does not factor in the fact that some

categories have a large variance, which can lead to misclassification

47


– Minimum distance classifiers– Parallelepiped classifiers– Maximum likelihood classifiers

48

Parallelepiped Classifiers

f

f

f

ff

Step 1 – define the range of values in each training area and use these ranges to construct an n-dimensional box (a parallelepiped) around each class

cc

c

cc

49


Step 2 – use the multi-dimensional ranges to create different surface categories

Fig 21

50


Regions do not have to be perfect rectangles, but can be “stepped”

This reduces errors

This approach does not classify pixels that fall outside of the defined areas

Fig 22

51


– Minimum distance classifiers– Parallelepiped classifiers– Maximum likelihood classifiers

52


Fig 23

53

Fig 24

54


Fig 25

55


Maximum likelihood classifiersEqui-probability contours define the level of statistical confidence you have in the classification accuracy

The smaller the contour, the higher the statistical confidence

Fig 26

56

Unsupervised Classification


Allow the computer to identify clusters based on different classification procedures

Fig 27

57

Hybrid Classification Approach

1. Perform an unsupervised classification to create a number of land cover categories within the area of interest

2. Carry out field surveys to identify the land cover type represented by different unsupervised clusters

3. Use a supervised approach to combine unsupervised clusters into similar land cover categories

58

Sources of Uncertainty in Image Classification

1. Non-representative training areas2. High variability in the spectral

signatures for a land cover class3. Mixed land cover within the pixel

area

59

Mixed pixels

In many cases, the IFOV of a sensor will include multiple land cover categories – e.g., a mixed pixel

Mixed pixels contribute to classification errors

Fig 27

60

ff

f

ff

cc

c

cc

dd

d

dd

m

m

m

m m

Question – How do different algorithms treat mixed pixels?

In some cases, mixed pixels are close enough to a specific category, which leads to misclassifications

61







62

Vegetation Indices

• Dimensionless, radiometric measures that serve as indicators of relative abundance and activity of green vegetation

63

Based on image presented at http://www.infocarto.es/vi.htm

Increasing leaf area

Red reflectance

64

Simple Vegetation Index (VI)

VI = RNIR / Rred

WhereRIR is the reflectance in the red band

Rredis the reflectance in the near infrared band

65

Problems with simple VI

• In areas with low vegetation, variations in the reflectance in the red channel from differences in the soil reflectance (from variations in soil moisture and different soil types) can cause changes in the VI independent of the vegetation cover

66

Normalized Vegetation Difference Index (NDVI)

RIR - Rred

NDVI = _________RIR +Rred

67

NDVI – relationship to intercepted NDVI – relationship to intercepted PARPAR

68

Why use vegetation indices?

• Partially normalize the external effects of sun angle, viewing angle, and atmospheric effects

• Partially normalize internal effects of shadowing, soil, amount of woody material, etc

• Can be directly linked to biophysical parameters, such as leaf area index, amount of green leaf biomass, amount of photosynthetic material

69

Sources of variation in satellite observed NDVI

• Differences in the overall level of vegetation cover– Broad categories of vegetation type –

savannas, shrubland, coniferous vs. deciduous forests

• Seasonal phenology – changes in green biomass associated with seasonal growth patterns and spring green up and fall senescence

• Inter-annual variations in climate• Disturbances that reduce green vegetation

– Deforestation, fire, insect outbreaks

70







72False-color SWIR composite image – areas in red are recent fires

73

Dall City Fire Black Spruce Stands

0

5

10

15

20

25

1 2 3 4 5 7

Landsat ETM Band

TO

A R

efle

ctan

ce (

%)

-

Unburned

Moderate Severity

High Severity

Extreme Severity

74

Normalized Burn Ratio (NBR)

• Using Landsat TM/ETM+ data

NBR = (B7-B4)/(B7+B4)

dNBR = (Post-burn NBR) – (Pre-burn NBR)

75

Composite Burn Index (CBI)

• CBI is a field measure of damage to a vegetated ecosystem by a fire

• Based on estimating levels of damage to specific characteristics in 5 strata – Overstory trees, Understory trees, Tall shrubs, Low shrubs/vegetation, Substrate

76

National Park Service• 286 ground plots sampled• Variety of vegetation types

and years sampled (pre- 2004 fires)

• R2 = 0.70 all fires combined

77

Data courtesy of J. Allen, NPS

78

dNBR maps of burn Severity

• Based on correlations between dNBR and CBI in different regions, a project has been started by USFS, USGS, and NPS to generate burn severity maps of all fires > 1,000 ha in the U.S. between 1984 and 2010 using Landsat TM/ETM+ data

• Called the Monitoring Trends in Burn Severity (MTBS) project

79







80

AVHRR 11 Sept 1999

Landsat ETM11 Sept 1999

81

AVHRR Vegetation Indices

• The availability of the long-term satellite data set from AVHRR (since 1978) spurred much interest in developing approaches to use these data to analyze global vegetation cover

• The NDVI is an effective approach to analyzing global vegetation cover– Simple, yet contains meaningful

information

82

Categories of AVHRR Data

• Local Area Coverage – LAC– Sampled at the full resolution (1.1 km)

of the AVHRR System– Requires downloading data at a ground

receiving station

• Global Area Coverage – GAC– Sub-sampled data with a 4 km resolution– Recorded onboard the satellite

83

GAC Sampling Protocol

Sampled pixel

Unsampled pixel

84

AVHRR GAC Data

• GAC data are resampled to create an effective pixel size of 4 by 4 km

85

AVHRR Composite Imagery

• While can AVHRR image the earth every day, it is not possible to obtain imagery of the earth every day because of cloud cover

• Studies have shown that during over a 1 to 2 week period, most of the earth’s surface is at sometime cloud free

• An NDVI composite image is one that takes the maximum NDVI over a specified period of time (1 week, 10 days, 2 weeks), and enters it into the data set

• In this way, global NDVI products are generated

86

Global Vegetation Index

• An AVHRR product created by NOAA• Re-sampled to a 16 km pixel size• All afternoon passes of AVHRR are

collected• A simple VI index (IR – red) is created

from all passes• Pixels with the highest VI over the entire

week are selected• NDVI calculated from the values in these

pixels

87

GLOBAL VEGETATION INDEX PRODUCT

• Products available from NOAA athttp://www.osdpd.noaa.gov/PSB/

IMAGES/ndvi_arc.html

90

Seasonal Variations in NDVI in an Alaskan Aspen Forest

0

0.2

0.4

0.6

0.8

120 150 180 210 240 270 300

Day of year

ND

VI

May Jun Jul Aug Sep Oct

91

Seasonal Variations in NDVI in Alaska Forests

0

0.2

0.4

0.6

0.8

120 150 180 210 240 270 300

Day of year

ND

VI

Black Spruce White Spruce Aspen

May Jun Jul Aug Sep Oct

92







93

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

1 2 3 4 5 6 7 8 9 10 11 12

Month of Year

ND

VI Growing season

length

Rate of greenup Rate of scenesence

Minimum NDVI

MaximumNDVI

NDVI Threshold

94From DeFries, Hansen, and Townshen, Global discrimination of land cover types from metrics derived from AVHRR pathfinder data, Remote Sensing of Environment 54:209-222, 1995.

95

Lecture Topics1. Goals for Image Processing2. Steps for producing an information product

from satellite data3. Definition of image classification4. Image slicing or thresholding




7. Analysis of temporal trends in vegetation indices

8. Decision tree classifiers

96

Decision Tree Classifier

• Decision tree classifiers use a simple set of rules to divide pixels into different land cover types

97

1: Evergreen Needleleaf Forests; 2: Evergreen Broadleaf Forests; 3: Deciduous Needleleaf Forests; 4: Deciduous Broadleaf Forests; 5: Mixed Forests; 6: Woodlands; 7: Wooded Grasslands/Shrubs; 8: Closed Bushlands or Shrublands; 9: Open Shrublands; 10: Grasses; 11: Croplands; 12: Bare; 13: Mosses and Lichens

http://www.geog.umd.edu/landcover/8km-map.html

98

Classification logic

a) a)

b)

b)

c)

c)

1 Lectures 11 Analysis and Classification of Digital MSS Data 21 October 2008 Today’s Lecture will...

Documents

Transcript of 1 Lectures 11 Analysis and Classification of Digital MSS Data 21 October 2008 Today’s Lecture will...