مواضيع المحاضرة: Intro of digital image processing
background image

Intro of digital image processing


background image

Remote Sensing Raster (Matrix) Data Format

Digital number of column 5, 

row 4 at band 2 is expressed 

as BV

5,4,2

= 105. 


background image

120 150 100 120 103

176 166 155

85 150

85

80

70

77 135

103

90

70 120 133

20

50

50

90

90

76

66

55

45 120

80

80

60

70 150

100

93

97 101 105

210 250 250 190 245

156 166 155 415 220

180 180 160 170 200

200

0 123 222 215

Band 2

Band 3

Band 4

1,1,2 2,1,2 3,1,2 4,1,2 5,1,2

1,2,2 2,2,2 3,2,2 4,2,2 5,2,2

1,3,2 2,3,2 3,3,2 4,3,2 5,3,2

1,4,2 2,4,2 3,4,2 4,4,2 5,4,2

Matrix notation for band 2

10

15

17

20

21

15

16

18

21

23

17

18

20

22

22

18

20

22

24

25

20

50

50

90

90

76

66

55

45 120

80

80

60

70 150

100

93

97 101 105

120 150 100 120 103

176 166 155

85 150

85

80

70

77 135

103

90

70 120 133

210 250 250 190 245

156 166 155 415 220

180 180 160 170 200

200

0 123 222 215

BIL

10

15

17

20

21

20

50

50

90

90

120 150 100 120 103

210 250 250 190 245

15

16

18

21

23

76

66

55

45 120

176 166 155

85 150

156 166 155 415 220

17

18

20

22

22

80

80

60

70 150

85

80

70

77 135

180 180 160 170 200

18

20

22

24

25

100

93

97 101 105

103

90

70 120 133

200

0 123 222 215

BSQ

10

20

120

210

15

15

76

176

156

16

17

80

85

180

18

18

100

103

200

20

50

150

250

17

50

66

166

166

18

55

80

80

180

20

60

93

90

0

22

97

100

250

20

90

120

155

155

21

45

85

70

160

22

70

77

70

123

24

101

120

190

21

90

103

245

415

23

120

150

220

170

22

150

135

200

222

25

105

133

215

BIP


background image

What is image processing

Is enhancing an image or extracting 

information or features from an image

Computerized routines for information 

extraction (eg, pattern recognition, 

classification) from remotely sensed 

images to obtain categories of information 

about specific features.

Many more


background image

Image Processing Includes 

Image quality and statistical evaluation

Radiometric correction

Geometric correction

Image enhancement and sharpening

Image classification

Pixel based

Object-oriented based

Accuracy assessment of classification

Post-classification and GIS

Change detection

GEO5083: Remote Sensing Image Processing and Analysis, spring 2013


background image

Image Quality

Many remote sensing datasets contain high-quality, 

accurate data. Unfortunately, sometimes error (or 

noise) is introduced into the remote sensor data by: 

the environment

(e.g., atmospheric scattering, 

cloud), 

random or systematic malfunction

of the remote 

sensing  system (e.g., an uncalibrated detector 

creates striping), or 

improper pre-processing

of the remote sensor 

data prior to actual data analysis (e.g., inaccurate 

analog-to-digital conversion). 


background image

155

154

155

160

162

163

164

MODIS
True
143

Cloud


background image

Clouds in ETM+


background image

Striping Noise and Removal

CPCA

Combined Principle 
Component Analysis

Xie et al. 2004


background image

Speckle Noise and 

Removal

G-MAP

Blurred objects
and boundary

Gamma Maximum 
A Posteriori Filter


background image

background image

Types of radiometric correction

Detector error or sensor error (internal 

error)

Atmospheric error (external error)

Topographic error (external error)


background image

Atmospheric correction

There are several ways to 

atmospherically correct 

remotely sensed data. 

Some are relatively 

straightforward while 

others are complex, 

being founded on 

physical principles and 

requiring a significant 

amount of information to 

function properly. This 

discussion will focus on 

two major types of 

atmospheric correction:

Absolute atmospheric 

correction

, and

Relative atmospheric 

correction

.

Solar  

irradiance 

Reflectance from  

study area,

Various Paths of  

Satelli te Received Radiance

Diffuse s ky  

irradiance 

Total radiance  

at the sensor

L

 L

 L 

Reflectance from  
 neighboring area,

1

2

3

Remote  

sens or 

detector

Atmosphere

5

4

1,3,5

E

L

90Þ



0

T



0

v

p

T

S

I



n

r



r

E

d

60 miles
or
100km

Scattering, Absorption
Refraction, Reflection


background image

Absolute atmospheric correction

Solar radiation is largely unaffected as it travels through the 

vacuum of space. When it interacts with the Earth’s atmosphere, 

however, it is selectively 

scattered and absorbed

. The sum of 

these two forms of energy loss is called 

atmospheric attenuation

.

Atmospheric attenuation may 1) make it difficult to relate hand-

held 

in situ

spectroradiometer measurements with remote 

measurements, 2) make it difficult to extend spectral signatures 

through space and time, and (3) have an impact on classification 

accuracy within a scene if atmospheric attenuation varies 

significantly throughout the image.

The general goal of 

absolute radiometric correction

is to turn 

the digital brightness values (or DN) recorded by a remote sensing 

system into 

scaled surface reflectance

values. These

values can 

then be compared or used in conjunction with scaled surface 

reflectance values obtained anywhere else on the planet.


background image

a) Image containing substantial haze prior to atmospheric correction. b) Image after 
atmospheric correction using ATCOR (Courtesy Leica Geosystems and DLR, the 
German Aerospace Centre). 


background image

Topographic correction

Topographic slope and aspect also introduce 

radiometric distortion (for example, areas in 

shadow)

The goal of a slope-aspect correction is to 

remove topographically induced illumination 

variation so that two objects having the same 

reflectance properties show the same 

brightness value (or DN) in the image despite 

their different orientation to the Sun’s position

Based on DEM, sun-elevation


background image

Conceptions of geometric correction

Geocoding:

geographical referencing

Registration:

geographically or nongeographically (no coordination system)

Image to Map (or Ground Geocorrection)

The correction of digital images to ground coordinates using ground control 

points collected from maps (Topographic map, DLG) or ground GPS points. 

Image to Image Geocorrection

Image to Image correction involves matching the coordinate systems or 

column and row systems of two digital images with one image acting as a 

reference image and the other as the image to be rectified.

Spatial interpolation:

from input position to output position or coordinates. 

RST (rotation, scale, and transformation), Polynomial, Triangulation

Root Mean Square Error (RMS): The RMS is the error term used to 

determine the accuracy of the transformation from one system to another. It is 

the difference between the desired output coordinate for a GCP and the actual.

Intensity (or pixel value) interpolation (also called resampling):

The process of 

extrapolating data values to a new grid, and is the step in rectifying an image that 

calculates pixel values for the rectified grid from the original data grid. 

Nearest neighbor, Bilinear, Cubic


background image

Image enhancement

image reduction, 

image magnification, 

transect extraction, 

contrast adjustments (linear and non-linear),

band ratioing, 

spatial filtering, 

fourier transformations, 

principle components analysis, 

texture transformations, and 

image sharpening


background image

Purposes of image classification

Land use and land cover (LULC)
Vegetation types
Geologic terrains
Hydrocarbon and Mineral exploration
Alteration mapping
…….


background image

What is image classification or 

pattern recognition

Is a process of classifying multispectral (hyperspectral) images into 

patterns of varying gray or assigned colors

that represent either 

clusters

of statistically different sets of multiband data, some of which 

can be correlated with separable classes/features/materials.  This is the 

result of 

Unsupervised Classification

, or 

numerical discriminators

composed of these sets of data that have been 

grouped and specified by associating each with a particular 

class

, etc. 

whose identity is known independently and which has representative 

areas (training sites) within the image where that class is located. This is 

the result of 

Supervised Classification

Spectral classes

are those that are inherent in the remote sensor 

data and must be identified and then labeled by the analyst.

Information classes

are those that human beings define. 


background image

supervised classification

Identify known a priori 

through a combination of fieldwork, map 
analysis, and personal experience as 

training 

sites

; the spectral characteristics of these sites are 

used to train the classification algorithm for 
eventual land-cover mapping of the remainder of 
the image. Every pixel both within and outside the 
training sites is then evaluated and assigned to the 
class of which it has the highest likelihood of 
being a member.

unsupervised classification

, The 

computer or algorithm automatically 
group pixels with similar spectral 
characteristics (means, standard 
deviations, covariance matrices, 
correlation matrices, etc.) into unique 
clusters according to some statistically 
determined criteria. The analyst then 
re-labels and combines the spectral 
clusters into information classes. 


background image

background image

Unsupervised classification

Uses 

statistical techniques

to group n-dimensional data into their natural 

spectral clusters, and uses the 

iterative procedures

label certain clusters as specific information classes

K-mean and ISODATA

For the first iteration arbitrary 

starting values

(i.e., the cluster properties) 

have to be selected. These 

initial values

can influence the outcome of the 

classification.

In general, both methods assign first arbitrary initial cluster values. The 

second step classifies each pixel to the closest cluster. In the third step the 

new cluster mean vectors are calculated based on all the pixels in one 

cluster. The second and third steps are repeated until the "change" between 

the iteration is small. The "change" can be defined in several different ways, 

either by measuring the distances of the mean cluster vector have changed 

from one iteration to another or by the percentage of pixels that have 

changed between iterations. 

The 

ISODATA algorithm has some further refinements

by splitting and 

merging of clusters. Clusters are merged if either the number of members 

(pixel) in a cluster is less than a certain threshold or if the centers of two 

clusters are closer than a certain threshold. Clusters are split into two 

different clusters if the cluster standard deviation exceeds a predefined value 

and the number of members (pixels) is twice the threshold for the minimum 

number of members.


background image

Supervised classification:

training sites selection 

Based on known a priori through a combination of fieldwork, 

map analysis, and personal experience

on-screen selection

of polygonal training data (ROI),

and/or 

on-screen seeding

of training data (ENVI does not have 

this, Erdas Imagine does). 

The 

seed program

begins at a single 

x, y 

location and evaluates 

neighboring pixel values in all bands of interest. Using criteria 

specified by the analyst, the seed algorithm expands outward like 

an amoeba as long as it finds pixels with spectral characteristics 

similar to the original seed pixel. This is a very effective way of 

collecting homogeneous training information. 

From 

spectral library

of field measurements


background image

Selecting

ROIs

Alfalfa

Cotton

Grass

Fallow


background image

Supervised classification methods

Various supervised classification algorithms may be used to assign an unknown pixel to one 

of 

m

possible classes. The choice of a particular classifier or decision rule depends on the 

nature of the input data and the desired output. 

Parametric

classification algorithms 

assumes that the observed measurement vectors 

X

c

obtained for each class in each spectral 

band during the training phase of the supervised classification are 

Gaussian

; that is, they are 

normally distributed. 

Nonparametric

classification algorithms make no such assumption. 

Several widely adopted nonparametric classification algorithms include:

one-dimensional 

density slicing

parallepiped

,

minimum distance

nearest-neighbor

, and 

neural network 

and

expert system analysis

.

The most widely adopted parametric classification algorithms is the:

maximum likelihood

.

Hyperspectral classification methods

Binary Encoding

Spectral Angle Mapper

Matched Filtering

Spectral Feature Fitting

Linear Spectral Unmixing


background image

Source: http://popo.jpl.nasa

.gov/html/data.html

Supervised

classification

method:

Spectral Feature

Fitting


background image

Accuracy assessment of classification

Remote sensing-derived thematic information are 

becoming increasingly important. Unfortunately, they 

contain errors.

Errors come from 5 sources:

Geometric error still there

None of atmospheric correction is perfect

Clusters incorrectly labeled after unsupervised classification

Training sites incorrectly labeled before supervised 

classification

None of classification method is perfect

We should identify the sources of the error, minimize it, 

do accuracy assessment, create metadata before being 

used in scientific investigations and policy decisions. 

We usually need GIS layers to assist our classification.


background image

Post-classification and GIS

salt-
and-
pepper


background image

types

Majority/Minority Analysis

Clump Classes

Morphology Filters

Sieve Classes

Combine Classes

Classification to vector (GIS)


background image

Change detection

Change detect involves the use of multi-temporal datasets to 

discriminate areas of land cover change between dates of imaging.

Ideally, it requires 

Same or similar sensor, resolution, viewing geometry, spectral bands, 

radiomatric resolution, acquisition time of data, and anniversary dates

Accurate spatial registration (less than 0.5 pixel error)

Methods

Independently classified and registered, then compare them

Classification of combined multi-temporal datasets, 

Principal components analysis of combined multi-temporal datasets

Image differencing (subtracting), (needs to find change/no change threshold, 

change area will be in the tails of the histogram distribution)

Image ratioing (dividing), (needs to find change/no change threshold, 

change area will be in the tails of the histogram distribution)

Change vector analysis

Delta transformation


background image

Example: stages of development


background image

1994

1996

Sun City –

Hilton Head


background image

1974

1,040 urban

hectares

1994

3,263 urban

hectares

315% 

increase




رفعت المحاضرة من قبل: Medoo Chan
المشاهدات: لقد قام 5 أعضاء و 373 زائراً بقراءة هذه المحاضرة








تسجيل دخول

أو
عبر الحساب الاعتيادي
الرجاء كتابة البريد الالكتروني بشكل صحيح
الرجاء كتابة كلمة المرور
لست عضواً في موقع محاضراتي؟
اضغط هنا للتسجيل