R sprojec3

9
Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina Page 1 Report on comparison of image classification on ENVI and Ecognition Prepared By: Sushmita Timilsina December, 2013 Department Of Civil and Geomatics Engineering (DOCGE)

Transcript of R sprojec3

Page 1: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 1

Report on comparison of image

classification on ENVI and Ecognition

Prepared By: Sushmita Timilsina

December, 2013

Department Of Civil and Geomatics Engineering (DOCGE)

Page 2: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 2

Contents Provided resources ................................................................................................................................. 3

Image classification in ENVI .................................................................................................................... 3

Image classification in ECOGNITION ....................................................................................................... 5

Final comparisons ................................................................................................................................... 8

Conclusion ............................................................................................................................................... 9

Page 3: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 3

Provided resources ENVI 4.3 software

Practical handbook

Ecognition software with user manual

Image file

Image classification in ENVI

The image provided to us was of a part of Chitwan District. The image was first subjected to

correction, filtering and then classified to a certain group.

At first the correction was applied to the image. Radiometric consistency among collocated

multi-temporal imagery is difficult to maintain, however, due to variations in sensor

characteristics, atmospheric conditions, solar angle, and sensor view angle that can obscure

surface change detection. For the correction we apply radiometric correction. Talking about

radiometric correction, there are two types of radiometric correction namely absolute and

relative radiometric correction. Absolute radiometric models use in situ measurements or

reasonable estimation of 64 X. Remote Sensing of Environment 98 (2005) 6379atmospheric

optical depth, solar zenith angle and satellite status to input parameters for calculating the

ground surface Reflectance. The goal of this study is to develop an improved PIF relative

radiometric normalization method for remote change detection, using multi-spectral imagery.

The goal of this study is to develop an improved PIF relative radiometric normalization

method for remote sensing. Since vegetation is one of our interested area for remote sensing

classification. One of the most popular vegetation indices is the normalised difference

vegetation index (NDVI). NDVI is usually expressed in the following form:

NDVI = (NIR - R) / (NIR + R)

ENVI allows us to calculate spectral indices using bands within a single image or using bands

from several images (as is the case for our data) using the band math function.

When the correction was made radiometric we then proceed to filtering of data. We

underwent with both high and low filter. Convolution filter are the filter that uses a moving

window for filtering and high pass filter are the filter that sharpens the appearance of fine

detail in an image. The moving window are in the form of matrix and the size of the matrix is

defined the user himself e.g 5* 5. The mean of all the pixels present in the matrix of the

original image is taken and is given to the new matrix formed.The high pass filter highlights

or sharpens the appearance of the fine detail in image so the finer details are seen through

high pass convolution.

Page 4: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 4

Low pass filter is designed to emphasis larger, homogeneous areas of similar tone and reduce

smaller detail in image. Being a convolution filter it also uses moving window average. The

moving window are in the form if matrix; size defined by the user e.g 3*3 or 5*5. In this the

standard deviation of the values of the pixels of the matrix of the original image is taken and

given to the matrix of the new filtered image. The low pass convolution filter filters the lesser

detail from the image. Hence the result is a smoothen image.

Page 5: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 5

When the correction and filtering was done to the image it was then subjected to the

classification. There were two types of classification made one with the corrected image and

one with the original one. We then compare it with the one classified in ENVI and one on the

ecognition.

Image classification in ECOGNITION Object-oriented image analysis is divided into three steps:

Multiresolution Segmentation

Create General Classes

Classification Rules

During the first step, image segments are defined and calculated. Parameters are defined by

the 2 user for the scale, spectral properties and shape properties. These image segments have

to be calculated on several hierarchical levels in a “trial and error” process to result in final

image segments to represent single objects of interest . The organization of the workflow is as

follows: 1) Input images, 2) Multiresolution segmentation, 3) Image object hierarchy, 4)

Creation of class hierarchy, 5) Classification using Training samples and standard nearest

neighbor, 6) Classification base Segmentation, 7) Repeat steps for best result, and 8) Final

merge classification. eCognition software is loaded with practice tutorials to understand the

basics of the eCognition software and its tools. The first tutorials show how to load and

display raster data, perform image segmentation, create a simple class hierarchy, insert the

nearest neighbor classifier into the class description, classify, and perform classification

quality assessment.

In ecognition we loaded the image given to us and then we did two procedures as follows

Chessboard

Multi-spectral

The image was loaded into eCognition and a spatial subset of the south eastern region of the

image was used. This area included a large water bodies, agriculture area, rural regions.Using

the “Layer Mixing Tool” in eCognition, the image parameters were defined by setting the

Equalizing to histogram and Preset to 3-layer mix.

Page 6: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 6

The second step is the multiresolution segmentation. The segmentation parameters were

defined. Layer weights were set to equal to one, scale equal to 20, shape factor to 0.3, and

color to 0.7, and compactness and smoothness to 0.5. The scale parameter is an abstract value

to determine the maximum possible change of heterogeneity caused by fusing several objects.

Color is the pixel value. Shape includes compactness and smoothness which are two

geometric features that can be used as "evidence." Smoothness describes the similarity

between the image object borders and a perfect square. Compactness describes the

"closeness" of pixels clustered in an object by comparing it to a circle.

Page 7: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 7

After defining the parameters, eCognition produces a new image with the new grouping of

pixels.

The next step includes creating Class Hierarchy by creating and defining classes. Three

classes were created: Agriculture, Rural, Wate. These 3 classes will be defined in the next

following steps. After creating the four classes, the tutorial defines the type of classifier

(nearest neighbor or membership functions). Nearest neighbor classifier is used by using the

Edit Standard Nearest Neighbor Feature Space Tool. For each class, the Standard Nearest

Neighbor Expression was inserted. In order to define the four classes, Training Sites of

known areas from the image can be created. With the specific class activated, training sites

can be selected by double clicking on the polygon. Image object data for the selected polygon

can be used. Nearest neighbor classification in eCognition is similar to supervised

classifications in common image analysis software. You have to declare training areas, which

are typical representatives of a class. In eCognition such training areas are referred to as

samples or sample objects. Multiple training sites for an each class were created. The

following color scheme for the training sites are:

Page 8: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 8

Green (Agriculture),

Blue (Water), Red

brown (Rural).

An accuracy assessment of was performed on the classification results. The best

classification result shows statistics of the training sites and classification description. These

statistics will allow you to compare which classes have the best classification based on the

preliminary results. A confusion matrix of the training sites with the classification was

produced. The overall accuracy is 100% because this statistics based on the samples, not the

final classification. A more accurate statistic can be produced by reclassifying the classes and

defining the class hierarchy with known field data. The known field data trainings sites will

reclassify the image and produce better statistics.

Final comparisons between the image classification between ENVI and Ecognition is

shown below:

Page 9: R sprojec3

Report on comparison of image classification on ENVI and Ecognition - Sushmita Timilsina

Page 9

Real image classification from ENVI classification in ecognition

Conclusion The ENVI has two classification procedure namely supervised and unsupervised. The

unsupervised has two way of performing from k-means and ISO data. The supervised

classification is more standard and expertised. It includes:

Knowledge Engineer: Definition of hypotheses, rules and variables

Knowledge Classifier: Rule-based classification

Iterative approach

Classification in ecognition is object orientated classification. Also the segmentation has gone

through for classification. It includes:

Hierarchical network of segments in different scale levels

Choice of proper thresholds for homogeneity and heterogeneity criteria

Trial & error