Image Classification R.A.ALAGU RAJA Remote Sensing & GIS Lab Department of ECE Thiagarajar College...

Post on 25-Dec-2015

218 views 4 download

Transcript of Image Classification R.A.ALAGU RAJA Remote Sensing & GIS Lab Department of ECE Thiagarajar College...

Image Classification

R.A.ALAGU RAJARemote Sensing & GIS Lab

Department of ECEThiagarajar College of Engineering

Madurai

Image Classification

• Aim is to automatically categorize all pixels in an image into land cover classes or themes.

Types• Supervised Classification• Unsupervised Classification

Scanner Measurement

Supervised Classification

• The image analyst “supervises” the pixel categorization process by specifying numerical descriptors of the various land cover types present in a scene – i.e. Requirement of Training areas

Steps:• Training Stage• Classification Stage• Output Stage

Supervised Classification - StepsTraining Stage :

• The image analyst identifies representative training areas and develops a numerical description of various land cover types.

Classification Stage :• Each pixel in the image data set is categorized into

the landcover class it most closely resembles.

Output Stage :• The output of classification will be in three typical

forms.• Thematic Maps• Tables• Digital Data files amenable to inclusion in a GIS.

Supervised Classification - Steps

Training Stage

Training Set Selection

Scatter Diagram

Minimum-Distance-to-Means • One of the simpler classification strategy.

• Mean Vector Formation.

• A pixel of unknown identity may be classified

by computing the distance between the value

of the unknown pixel and each of the category

means.

• If the pixel is farther than an analyst – defined

distance from any category mean – unknown.

Minimum-Distance-to-Means

Minimum Distance to Mean Classification

Parallelepiped Classification

• Sensitivity to category variance is introduced

• The range is defined by the highest and lowest DN values – Rectangular Area.

• Parallelepipeds – The multidimensional analogs of the rectangular areas.

• Very fast and computationally efficient.

• When category ranges overlap – Difficulties are encountered.

• Unknown pixels – Classified as not sure (or) arbitrarily placed in any one (or both) of the two overlapping classes.

Parallelepiped Classification

Stepped Border Parallelepiped

• Covariance – Tendency of spectral values to vary

similarly in two bands – “Slanted Clouds of

Observations”.

• Corn & Hay Category – Exhibits positive covariance.

• Water Category – Exhibits Negative Covariance.

• In the presence of covariance, the rectangular

decision regions fit the category training data very

poorly.

• Solution – Modifying the single rectangles into a

series of rectangles with stepped borders.

Stepped Border Parallelepiped

Parallelepiped classification

Gaussian Maximum Likelihood Classifier• The MLC quantitatively evaluates both the

variance and covariance of the category spectral

response patterns.

• The algorithm calculates the probability of an

unknown pixel being a member in each category.

• The pixel is assigned in the most likely class

(Highest probability values).

Probability Density Function

Maximum Likelihood Classifier

Maximum Likelihood Classification

Maximum Likelihood Classification Report

Sl. No.

Themes Pixel 1992 Area 1992 (Sq.Km)

Pixel 1997 Area 1997 (Sq.Km)

1. Settlement 1436948 1888.239 1566874 2058.970

2. Water 15271 20.067 10567 13.885

3. Hills 730164 959.481 722925 949.968

4. Unused Lands 675868 888.132 630789 828.896

5. Vegetation 718914 944.697 720931 947.348

6. Background 2267835 2980.076 2252914 2960.469

7. Null 0 0 0 0

Total 5875000 7720.114 5875000 7720.114

Case StudyUrban Sprawl Monitoring for

Madurai City Using

Multispectral Data Analysis

Urban Sprawl

• Urban sprawl is unplanned, uncontrolled

spreading of urban development into areas

adjoining the edge of a city.

• Urban sprawl leads to absence of regional

planning.

• Urban sprawl can be resolved by Remote

Sensing and Change Detection algorithms.

Objective

• To assess the urban growth by using

various change detection algorithms.

• To recommend an optimal change

detection algorithm for urban growth

monitoring.

Study Area

• Madurai City, Tamilnadu

• Referred as Athens of Asia

• Second Largest City in Tamil Nadu

• One of the Mini Metros (20 Cities) in India – Population

14,33,251 (Acc. Census 2001)

• Historical City with Rich Cultural Heritage

• Established in 7th Century A.D.

• Hot Tourist Destination

• Latitude : 90 50’ 59” N to 90 57’ 36” N

• Longitude : 780 04’ 47” E to 780 11’ 23” E

Satellite Image Processing:

It involves the manipulation and interpretation of satellite images with the aid of computers.

Classification: To automatically categorize all pixels in an image into

land cover classes or themes.

Change Detection: It is process of identifying differences in the state of an

object or phenomenon by observing it at different times.

Work Flow

Image Enhancement

Geometric correction

Resampling to 30 meter

Ground Truth Verification

Image Enhancement

Geometric correction

Resampling to 30 meter

Change Detection Algorithms

Image 1

Urban Sprawl Map

Image 2

Band Separator

(ID, IR, CVA)

Band Separator

(ID, IR, CVA)

Image Classifier

Image Classifier

Principal Component Analysis

Dataset

Image 1 Image 2

Satellite : IRS 1B IRS P6

Sensor : LISS II LISS III

Resolution : 36.25m 23.50m

Date : 4th Mar 96 19th Mar 04

Area of coverage : 148.5 sq.km

Enhanced Images

1996 Enhanced Image. 2004 Enhanced Image.

Image 2Image 1

Change Detection Algorithms

• Image Differencing

• Image Ratioing

• Post Classification Comparison

• Change Vector Analysis

• Principal Component Comparison

Image Differencing

• The most common technique to detect

changes of an image.

• Each pixel from an image is subtracted

from corresponding pixel in another image.

I.D =t2 – t1

Thresholding:

Chosen based on standard deviation value from the histogram plot.

Image Differencing

Image Date 2Difference Image = Image 1 - Image 2

Image Date 1

Image Differencing

Threshold image (ID).

Increased

Decreased

No change

Legend

Standard Deviation = 49

Tabulation of Result (ID)

Type of Change Pixels Area (sq. km)

No Change (Black) 155740 140

Increased radiance (Green)

4785 4.3

Decreased radiance (Red)

4711 4.2

Change and No change area (ID).

Total Area 148.5

• Comparison of two independently

classified images.

• Compute the Error matrix.

Post Classification Comparison

Classification

Maximum Likelihood Algorithm:

• Creates N-dimensional ellipsoids.

• Probability density function is calculated

for each pixel with respect to training data

sets.

• The pixel is classified into a type which

has maximum probability.

Classified Images

1996 Classified image. 2004 Classified image.

Scrub & ForestLand without Scrub

Already Builtup LandBuiltup Land

Tank

Wet Land

Scrub & Forest

Legend

PCC change map

Change map (PCC).

Already Builtup LandBuiltup Land

Land without ScrubTank

Wet Land

Scrub & Forest

Legend

Field Visit – Collection of GCPs Using GPS Receiver

S. No.

Name of the area

Latitude

Longitude Elevatio

nAccuracy Features

1.Kudhal Nagar tank

9.9511 78.1041 138 24 Vegetation

2. Sellur tank 9.9407 78.1184 148 27 Water body

3. SITCO 9.9415 78.1494 135 28 Urban

4. Ring road 9.8565 78.1196 127 22 Waste land

5. Chinthamani 9.8874 78.1438 133 24 Urban

Applications

• City Planning

• Mapping

• Population Estimation

• Site Selection

• Traffic Management and Parking studies

• Encroachment