GI 225: REMOTE SENSING APPLICATIONS

2020/2021

LECTURE 3: DIGITAL IMAGE ANALYSIS

• Introduction
• Feature extraction
• Classification

1
INTRODUCTION

• Image analysis is the extraction of meaningful information from

images; mainly from digital images by means of digital image
processing techniques
• Image analysis partition remote sensing imagery into
meaningful image-objects, and assessing their characteristics
through spatial, spectral and temporal scale
• Computers are indispensable for the analysis of large amounts
of data, for tasks that require complex computation, or for the
extraction of quantitative information.

2
INTRODUCTION (cont’d)
• On the other hand, the human visual cortex is an excellent
image analysis apparatus, especially for extracting higher-level
information, and for many applications including medicine,
security, and remote sensing
• Human analysts still cannot be replaced by computers.
• For this reason, many important image analysis tools such as
edge detectors and neural networks are inspired by human
visual perception models
• Computer image analysis largely contains the fields of computer
or machine vision, and makes heavy use of pattern recognition,
digital geometry, and signal processing. 3
INTRODUCTION (cont’d)
• Image analysis closely relates to image patter recognition
because they both cover similar fields.
• Like pattern recognition, image analysis also involves 4 steps:

Image Image Feature

Classification Decision
acquisition Processing extraction

• The first two steps have been covered, this lecture will focus on:

− Feature extraction

− Classification
4
FEATURE EXTRACTION
• The Feature extraction is a part of the dimensionality reduction
process, in which, an initial set of the raw data is divided and
reduced to more manageable groups.
• Feature extraction is important because large data sets have a
large number of variables.
• Analysis with a large number of variables generally requires a
large amount of memory and computation power
• Feature extraction constructs few combinations of the variables
to get around large dataset while still describing the data with
sufficient accuracy.
5
FEATURE EXTRACTION (cont’d)
• Some of the statistical or machine learning techniques for
image feature extraction includes:
− Isomap
− Partial least squares
− Latent semantic analysis
− Principal component analysis
− Independent component analysis
− Multifactor dimensionality reduction

6
FEATURE EXTRACTION (cont’d)

7
CLASSIFICATION
• Image classification - is the process of sorting pixels into a finite
number of individual classes or categories of data based on
their values.
• If a pixel satisfies a certain set of criteria it is assigned to the
class correspond to that criteria.

Unclassified Image Classified Image

8
CLASSIFICATION (cont’d)
• It converts multiband satellite image into single band image
• with different classes, defined or undefined.
• It groups similar features and separate dissimilar ones
resulting to a manageable size of classes

Band 1

Classification
Band 4

Band 3
Classified image
9
CLASSIFICATION (cont’d)
• Why image classification ?
− It is more objective than visual interpretation
− To find anomalous patterns in the image data set
− It is effective analysis of complex multi-band
interrelationships
− It is more cost efficient in the analyses of large datasets
− To translate continuous variability of image into map
patterns or thematic maps that provide meaning to the user
− To obtain insight in the data with respect to ground cover
and surface characteristics
− To produce land use/land cover map 10
CLASSIFICATION (cont’d)
• Commonly used methods for the classification of images are:
− Supervised classification
− Unsupervised classification

• In this lecture image classification will cover the following:

− Supervised classification
− Unsupervised classification
− Accuracy assessment
− Classification tools
− Applications
11
CLASSIFICATION (cont’d)
Supervised classification
• Is the process of using training samples of known identity to
classify pixels of unknown identity.
• It is based on how close a pixel to be classified is to each
training sample

• This classification
approach is based on:
− spectral groupings
− prior knowledge
− user interaction
12
CLASSIFICATION (cont’d)
Supervised classification
• Procedure involves:
1) Load unclassified image into the classification tool
2) Creates training samples to represent the classes of
interest
3) Classify using an appropriate classifier
4) Assess the accuracy of a classified image

• Supervised classification requires the analyst to select training

areas where he/she knows what is on the ground
13
CLASSIFICATION (cont’d)
Supervised classification
Analyst selects training
areas and digitize a polygon Mean Spectral
within that area… Then computer creates... Signatures
Conifer

Known Conifer
Area

Water
Known Water
Area

Deciduous
Known Deciduous
Area
Unclassified image
CLASSIFICATION (cont’d)
Supervised classification

Mean Spectral Information

Signatures Multispectral Image (Classified Image)

Conifer

Deciduous

Water Unknown

Next Pixel to
be Classified
CLASSIFICATION (cont’d)
Supervised classification

Land Cover Map

Legend
Water
Conifer
Deciduous
CLASSIFICATION (cont’d)
Supervised classification
• The choice of particular classifier depends on nature of input
data and output required

• Common supervised classification algorithms/ classifiers are:

− Parallelepiped
− Minimum-distance
− Maximum likelihood

17
CLASSIFICATION (cont’d)
Supervised classification
• Advantages:
− Useful information categories are distinct first
− Spectral separability is examined by the analyst

18
CLASSIFICATION (cont’d)
Supervised classification
• Disadvantages:
− Difficulty in selecting training sites because of the
variability of spectral response within each class.
− Time consuming in data collection
− It is subjective because spectral classes can be
distinguished and be revealed by the unsupervised
approach
− Depends heavily on the prior knowledge of the area
− Skills of the individual processing the image
− Distinctness of the classes
19
CLASSIFICATION (cont’d)
Unsupervised classification
• Is the process where a computer program automatically
groups the pixels into separate clusters based on their
spectral features
• It is an automatic process without the user interaction in
providing sample classes
• Computer select means and covariance matrices to be used
in classification
• Computer groups pixels with similar spectral characteristics
into unique clusters according to statistical criteria
20
CLASSIFICATION (cont’d)
Unsupervised classification
• Analyst attempts a posteriori to assign labels the spectral
clusters into information classes

• Some clusters may be meaningless because they represent

mixed classes 21
CLASSIFICATION (cont’d)
Unsupervised classification
• Things to be done by an analyst:
1) Pre-defines the number of iterations
2) Selects the number of clusters desired
3) Selects the minimum or maximum number of clusters in
process
• Things done by a computer algorithm:
1) Determines that pixels are related and groups them into
classes
2) Automatically assign each cluster a land cover type
• The final output is a "thematic map". 22
CLASSIFICATION (cont’d)
Unsupervised classification
The analyst requests the computer
to examine the image and extract a
Spectrally distinct clusters
number of spectrally distinct clusters… Cluster 3
Cluster 6

Cluster 5
Cluster 2

Cluster 1
Cluster 4

Unclassified image
 CLASSIFICATION (cont’d)
Unsupervised classification
Output Classified
Saved clusters Multispectral Image Image
Cluster 3
Cluster 6

Cluster 5
Cluster 2

Cluster 1
Cluster 4

Next Pixel to
be Classified
CLASSIFICATION (cont’d)
Unsupervised classification
• Clustering algorithm used for the unsupervised classification
generally vary according to the efficiency with which the
clustering takes place
• Common unsupervised classification algorithms/ classifiers
are:
− Chain method
− Isodata clustering

25
CLASSIFICATION (cont’d)
Unsupervised classification
• Advantages:
− It reveals discriminable classes
− Do not utilize training data as the basis for classification.
− It performs well in areas where the values of a given cover
type are close together in the measurement space,
− Data in different classes are comparatively well separated.

26
CLASSIFICATION (cont’d)
Unsupervised classification
• Disadvantages:
− Spectral Classes do not always correspond to information
Classes
− The user also has to spend time interpreting and label the
classes following the classification
− Spectral properties of classes can also change over time, so
you can’t always use same class when moving from one
image to another.

27
CLASSIFICATION (cont’d)
Accuracy assessment
• Is the comparison of the classified image to geographical data
that are assumed to be true, in order to determine the
accuracy of the classification process.
• It is performed by comparing the map created by
classification to a reference map based on a different
information source.
• Usually, the reference map are derived from ground truth
data.
• Primary purposes is to perform quantitative comparisons
of different interpretations. 28
CLASSIFICATION (cont’d)
Accuracy assessment
• Primary purposes is to perform quantitative comparisons
of different interpretations.
• In order to compare, both classified map and the reference
map must be accurately geometrically corrected

• Steps to asses accuracy

− Collect reference data
− Determine class types at specific locations
− Compare reference to classified map
29
CLASSIFICATION (cont’d)
Accuracy assessment
• Prior to accuracy assessment a contingency table or
error/confusion matrix is developed using samples from
classified and reference map.
• Consider an Example 1,

30
CLASSIFICATION (cont’d)
Accuracy assessment
• Diagonal elements represent pixels which are correctly
classified according to the reference data
• Off-diagonal elements represent misclassified pixels
• Quantification of the accuracy of classified image can be done
using the following:
− Overall accuracy (measures errors of omission)
− User accuracy (measures proportion of correct classified class)
− Producer accuracy (measures proportion of correct ground
truth in a class)
− Kappa (measures proportion of improvement by classifier)31
CLASSIFICATION (cont’d)
Accuracy assessment
• Quantification of accuracy (cont’d):
− Overall accuracy of classification (OAC) is the ratio of the
total number of correctly classified pixels to the total
number of all classified pixels in an image.

𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑙𝑙 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑝𝑖𝑥𝑒𝑙𝑠

𝑂𝐴𝐶 =
𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑙𝑙 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑝𝑖𝑥𝑒𝑙𝑠

− From confusion matrix in Example 1 above,

𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎𝑙𝑙 𝑐𝑙𝑎𝑠𝑠𝑒𝑠
𝑂𝐴𝐶 =
𝑐𝑜𝑙𝑢𝑚𝑛 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎𝑙𝑙 𝑐𝑙𝑎𝑠𝑠𝑒𝑠

(35 + 37 + 41)
𝑂𝐴𝐶 = = 83%
(39 + 50 + 47)
32
CLASSIFICATION (cont’d)
Accuracy assessment
• Quantification of accuracy (cont’d):
− User accuracy of classification (UAC) is the ratio of number
of correctly classified pixels of a class to the total number of
classified pixels in the same class.
𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠
𝑈𝐴𝐶 =
𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠

− From confusion matrix in Example 1 above,

𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠
𝑈𝐴𝐶 =
𝑟𝑜𝑤 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠

For class A, U𝐴𝐶 = 35 = 90%

39
33
CLASSIFICATION (cont’d)
Accuracy assessment
• Quantification of accuracy (cont’d):
− Producer accuracy of classification (PAC) is the ratio of
number of correctly classified pixels of a class to the total
number of ground truth pixels in the same class.
𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠
𝑃𝐴𝐶 =
𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑔𝑟𝑜𝑢𝑛𝑑 𝑡𝑟𝑢𝑡ℎ 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠

− From confusion matrix in Example 1 above,

𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠
𝑃𝐴𝐶 =
𝑐𝑜𝑙𝑢𝑚𝑛 𝑝𝑖𝑥𝑒𝑙𝑠 𝑜𝑓 𝑎 𝑐𝑙𝑎𝑠𝑠

For class A, 35
𝑃𝐴𝐶 = = 70%
50 34
CLASSIFICATION (cont’d)
Accuracy assessment
• Quantification of accuracy (cont’d):
− Kappa coefficient ෡)
(𝐾 is the measure of the proportional
improvement by the classifier over purely random
assignment to classes.
− Mathematically,

𝑖=𝑗=𝑛 𝑖=𝑗=𝑛
𝑁 × σ𝑖=𝑗=1 𝑑𝑖,𝑗 − σ𝑖=𝑗=1 𝑟𝑡,𝑗 × 𝑐𝑖,𝑡
෡=
𝐾 𝑖=𝑗=𝑛
𝑁 2 − σ𝑖=𝑗=1 𝑟𝑡,𝑗 × 𝑐𝑖,𝑡

Where, N is the number of all classified/ ground pixels

d sums diagonal pixels of all classes
r and c sum product of row total and column total.
35
CLASSIFICATION (cont’d)
Accuracy assessment
• Quantification of accuracy (cont’d):
− Kappa coefficient ෡)
(𝐾

− From confusion matrix Example 1 above

136 × (35 + 37 + 41) − (39 × 50 + 50 × 40 + 47 × 46)

෡=
𝐾
1362 − (39 × 50 + 50 × 40 + 47 × 46)

136 × 113 − 6112

෡=
𝐾
1362 − 6112

෡ = 75%
𝐾

36
CLASSIFICATION (cont’d)
Classification tools
• Some of the most common classification GIU software are:
− ArcGIS
− ILWIS
− IDRISI Andes
− ERDAS Imagine

• The most common classification programming tools are:

− Python
− R
− MATLAB
− Etc. 37
CLASSIFICATION (cont’d)
Applications
• The most common application of image classification is
production of land cover maps
• Land cover not necessarily equivalent to land use
− We focus on What’s there: Land cover
− Many users are interested in how what’s there is being
used: Land use
− Grass is land cover; pasture and recreational parks are land
uses of grass

38
ASSIGNMENTS
• Assignment 3.1: Briefly outline advantages and
disadvantages of a classification algorithm below as they
apply in image classification.
− Maximum likelihood – group 1 (HIP, URP, RDP)
− Isodata clustering – group 2 (HIP, URP, RDP)
− Parallelepiped – group 3 (HIP, URP, RDP)
− Chain method – group 4 (HIP, URP, RDP)
− Minimum-distance - group 2 (HIP, URP, RDP)

39
End

40