3.

Information Extraction Process

Prof. Dr-Ing. F.N. Karanja

 Outline
• 3.0. Introduction
• 3.1. Problem Definition
• 3.2. Identification of Image Data
• 3.3. Identification of the algorithm
• 3.4. Parameters Setting
• 3.5. Information Extraction
 Lecture Objectives

• At the end of this topic the learner should be able

to:-
– Define a remote sensing problem
– Identify suitable remote sensing data based on the
application
– Select the most appropriate algorithm
– Set and justify the parameters appropriately
– Extract the required information
 3.0 Introduction (1)
• Recall the Remote Sensing Process
Introduction (2)

Classes

DN Values
Introduction (4)
 Introduction (5)
• Remotely sensed data can be used to extract thematic and metric
information, making it ready for integrated processing.
• Thematic information provides descriptive data about earth surface
features.
• Themes can be as diversified as their areas of interest, such as soil,
vegetation, water depth, and land cover
• Metric information includes location, height, and their derivatives, such as
area, volume, slope angle, and so on.
• Thematic information can be obtained through visual interpretation of
remote sensing images (including photographs) or computer-based digital
image analysis.
• Metric information is extracted by using the principles of photogrammetry
which are also applicable to current digital image processing.
Information Extraction Process
Problem Definition
• The Problem should be
clear
• Definition of the
objectives to be achieved
should be done concisely,
i.e. they should be
SMART;
– S-Specific
– M-Measurable
– A-Achievable
– R-Realistic
– T-Time bound
• The end result should be
in line with the
objectives set.
 3.2. Identification of Image Data
 Depending on the application, the information classes of interest
need to be defined and their characteristics assessed. On the basis
of these, the most appropriate data can be selected and the criteria
include:-
 Sensor type (resolution, level of detail, etc) –(Spatial Resolution)
 Relevant wavelength bands (Spectral Resolution)
 Date of acquisition (important for phenomena that are time dependent, also
considerations in regard to cloud cover and amount of illumination should be
taken into account.- (Temporal Resolution)
 Project Time Frame
 Available Human capacity and skills
 Budgetary Criteria.
 Image Interpretation
• Interpretation is defined as the act of examining images for the purpose of identifying
objects and judging their significance (Colwell, 1997).
• The activities of aerial photo/image interpreters may include
– (1) detection/identification,
– (2) measurement, and
– (3) problem solving.
• In the process of detection and identification the interpreter identifies objects,
features, phenomena, and processes in the photograph and conveys his or her
response by labelling.
• These labels are often expressed in qualitative terms, for example, likely, possible,
probable, or certain.
• The interpreter may need to make quantitative measurements.
• At the stage of problem solving, the interpreter identifies objects from a
study of associated objects or complexes of objects from an analysis of their
component objects, and this also may involve examining the effect of some process
and suggesting a possible cause.
Commonly used interpretation Elements (1)
 Commonly used interpretation Elements (2)
• In the context of digital analysis of remotely sensed data, the basic elements of
image interpretation, although developed initially based on aerial photographs,
are applicable to digital images.
• However, most digital image analysis methods are based on tone or color,
which is represented as a digital number (i.e., brightness value) in each pixel of
the digital image.
• As multisensor and high spatial resolution data have become available, texture
has been used in image classification, as well as contextual information, which
describes the association of neighbouring pixel values.
• Before main image analyses take place, pre-processing of digital images often
is required.
• Image pre-processing may include detection and restoration of bad lines,
geometric rectification or image registration, radiometric calibration and
atmospheric correction, and topographic correction.
Spectral and Information Classes
• Image classification uses spectral
information represented by digital
numbers in one or more spectral
bands and attempts to classify each
individual pixel based on the spectral
information. The objective is to
assign all pixels in the image to
particular classes or themes (e.g.,
water, forest, residential, commercial,
etc.) and to generate a thematic
“map.”

• It is important to differentiate between

information classes and spectral
classes. The former refers to the
categories of interest that the analyst
is actually trying to identify from the
imagery, and the latter refers to the
groups of pixels that are uniform (or
near alike) with respect to their
brightness values in the different
spectral channels of the data.
 Data Calibration (1)
Generally, there are two approaches to sample definition to
image classification:
• In a supervised classification, the analyst identifies in the
imagery homogeneous representative samples of different
cover types (i.e., information classes) of interest to be used as
training areas. Each pixel in the imagery then would be
compared spectrally with the training samples to determine to
which information class they should belong. (Class notes of
DIP).
• In an unsupervised classification, spectral classes are first
grouped based solely on digital numbers in the imagery,
which then are matched by the analyst to information classes.
(Class notes DIP)
Data Calibration (2)
 3.3. Identification of Algorithm

• In recent years, many advanced classification

approaches, such as artificial neural network, fuzzy-
set, and expert systems, have become widely
applied for image classification including machine
and deep learning
 Per-Pixel Algorithms (1)
• Most classification approaches are based on per-pixel information, in which
each pixel is classified into one category, and thematic classes are mutually
exclusive.
• Traditional perpixel classifiers typically develop a signature by combining the
spectra of all training-set pixels for a given feature.
• The resulting signature contains the contributions of all materials present in
the training pixels.
• Per-pixel-based classification algorithms may be parametric or
nonparametric.
– The parametric classifiers assume that a normally distributed dataset exists and
that statistical parameters (e.g., mean vector and covariance matrix) generated
from the training samples are representative). However, the assumption of
normal spectral distribution is often violated, especially with complex landscapes
such as urban areas.
In addition, insufficient, nonrepresentative, or multimode distributed training
samples can introduce further uncertainty in the image classification procedure.
Another major drawback of the parametric classifiers lies in the difficulty in
integrating spectral data with ancillary data.
Per-Pixel Algorithms (2)
• With nonparametric classifiers, the assumption of a normal
distribution of the dataset is not required. No statistical
parameters are needed to generate thematic classes.
Nonparametric classifiers thus are suitable for the incorporation
of non-spectral data into a classification procedure.
• Previous research has indicated that nonparametric classifiers
may provide better classification results than parametric
classifiers in complex landscapes
• Among commonly used nonparametric classification methods
are neural-network, decision-tree, support-vector machine, and
expert systems.
• Bagging, boosting, or a hybrid of both techniques may be used
to improve classification performance in a nonparametric
classification procedure. These techniques have been used in
decision-tree and support-vector machine algorithms to
enhance image classification.
Per-Pixel Algorithms (3)
 Examples of Per Pixel Algorithms (1)
• Neural network
• Decision-tree classifier
• Spectral-angle classifier
• Supervised iterative
• classification (multistage classification)
• Enhancement classification approach
• Multiple-forward-mode (MFM-5-scale) approach to running the 5-scale
geometric optical reflectance model
• Iterative partially supervised classification based on a combined use of a
radial basis function network and a Markov random-field approach
• Classification by progressive generalization
• Support-vector machine
• Unsupervised classification based on independent components analysis
mixture model
 Examples of Per Pixel Algorithms (2)

• Optimal iterative unsupervised classification

• Model-based unsupervised classification
• Linear constrained discriminant analysis
• Multispectral classification based on
probability-density functions
• Layered classification
• Nearest-neighbor classification
• Selected-pixels classification
Example of a neural network process 1
Example of a neural network process 2
 Training and testing

Data acquisition Practical usage

Universal
set
(unobserved)

Training set Testing set

(observed) (unobserved)
 Sub-Pixel Algorithms (1)
• Owing to the heterogeneity of landscapes (particularly urban landscapes)
and the limitation in spatial resolution of remote sensing imagery, mixed
pixels are common in medium-and coarse-spatial-resolution data. The
presence of mixed pixels has been recognized as a major problem that
affects the effective use of remotely sensed data in per-pixel-based
classifications
• Subpixel-based classification approaches have been developed to provide
a more appropriate representation and accurate area estimation of land
covers within the pixels, especially when coarse-spatial-resolution data are
used.
• A fuzzy representation, in which each location is decomposed of multiple
and partial memberships of all candidate classes, is needed. Different
approaches have been used to derive a soft classifier, including fuzzy-set
theory, Dempster-Shafer theory, certainty factor softening the output of a
hard classification from maximum likelihood and neural networks
 Sub-Pixel Algorithms (2)
• In addition to the fuzzy image
classifier, other subpixel
mapping approaches also have
been applied.
• Among these approaches, the
fuzzy-set technique, ERDAS
IMAGINE’s subpixel classifier
and spectral mixture analysis
(SMA)–based classification are
the three most popular
approaches used to overcome
the mixed-pixel problem.
• An important issue for subpixel-
based classifications lies in the
difficulty in assessing
classification accuracy.
 Examples of Sub-Pixel Algorithms
• Imagine Subpixel Classifier
• Fuzzy classifier
• Fuzzy expert system
• Fuzzy neural network
• Fuzzy-based multisensor data fusion classifier
• Rule-based machine version approach
• Linear regression or linear least squares
inversion
 Per-Field Algorithms
• The heterogeneity in complex landscapes, especially in urban areas, results in high
spectral variation within the same land cover class.
• With per-pixel classifiers, each pixel is individually grouped into a certain category,
but the results may be noisy owing to high spatial frequency in the landscape.
• The per-field classifier is designed to deal with the problem of landscape
heterogeneity and has been shown to be effective in improving classification
accuracy
• A per-field-based classifier averages out the noise by using land parcels (called
fields) as individual units
• GIS provides a means for implementing per-field classification through integration
of vector and raster data .
• The vector data are used to subdivide an image into parcels, and classification then
is conducted based on the parcels, thus avoiding intra-class spectral variations.
• However, per-field classifications are often affected by such factors as the spectral
and spatial properties of remotely sensed data, the size and shape of the fields, the
definition of field boundaries, and land-cover classes chosen
 Examples of Per-Field Algorithms
• Per-field or per-parcel classification
• Per-field classification based on per-pixel or subpixel
classified image
• Parcel-based approach with two stages: per parcel
classification using conventional statistical classifier
and then knowledge-based correction using
contextual information
• Map-guided classification
• Object-oriented classification
• Graph-based structural pattern recognition system
• Spectral shape classifier
Example of a Per field classified image
Comparison of per pixel and field
Classifier
 Contextual Based Approaches

• Contextual classifiers have been developed to cope with the problem of

intraclass spectral variations in addition to object-oriented and per-field
classifications.
• Contextual classification exploits spatial information among neighboring
pixels to improve classification results.
• Contextual classifiers may be based on smoothing techniques, Markov
random fields, spatial statistics, fuzzy logic, segmentation, or neural
networks
• In general, presmoothing classifiers incorporate contextual information as
additional bands, and a classification then is conducted using normal
spectral classifiers, whereas postsmoothing classifiers use classified
images that are developed previously using spectral-based classifiers.
• The Markov random-fieldbased contextual classifiers such as iterated
conditional modes are the most frequently used approach in contextual
classification and have proven to be effective in improving classification
results.
Examples Contextual Based Approaches
• Extraction and classification of homogeneous objects (ECHO)
• Supervised relaxation classifier
• Frequency-based contextual classifier
• Contextual classification approaches for high-and low-resolution data,
respectively, and a combination of both approaches
• Contextual classifier based on region-growth algorithm
• Fuzzy contextual classification
• Iterated conditional modes
• Sequential maximum a posteriori classification
• Point-to-point contextual correction
• Hierarchical maximum a posteriori classifier
• Variogram texture classification
• Hybrid approach incorporating contextual information with per-pixel
classification
• Two-stage segmentation procedure
 Knowledge-Based Algorithms
• Evidential reasoning classification
• Knowledge-based classification
• Rule-based syntactical approach
• Visual fuzzy classification based on use of
exploratory and interactive visualization techniques
• Decision fusion–based multitemporal classification
• Supervised classification with ongoing learning
capability based on nearest-neighbor rule
Example of a knowledge based
classifiction
 Combination of multiple classifiers
approaches (1)
• Multiple classifier system (BAGFS: combines bootstrap aggregating with
multiple feature subsets)
• A consensus builder to adjust classification output (MLC, expert system,
and neural network)
• Integrated expert system and neural network classifier
• Improved neuro-fuzzy image classification system
• Spectral and contextual classifiers
• Mixed contextual and per-pixel classification
• Combination of iterated contextual probability classifier and MLC
• Combination of neural network and statistical consensus theoretical
classifiers
• Combination of MLC and neural network using Bayesian techniques
• Combining multiple classifiers based on product rule, staked regression
• Combined spectral classifiers and GIS rule-based classification
 Combination of multiple classifiers
approaches (2)

• Combination of MLC and decision-tree classifier

• Combination of nonparametric classifiers (neural network,
decision treeclassifier, and evidential reasoning)
• Combined supervised and unsupervised classification
3.4.Practical Example
Unsupervised and Supervised
Classification
Erdas Imagine Approach
Unsupervised: ( Algorithm: ISODATA)
Control Parameters: - Number of iterations
- Convergence limit
- Number of Classes
- Skip Factors (X,Y)
Labelling of Classes using the Raster Attribute
Editor
Example: SPOT XS
Unsupervised Classification
Parameters:
Number of iterations = 20; Convergence limit=0.95; Number of classes=10;
Skip Factors X=1; Y=1

Original Image Unsupervised Classified Image

Forest Agricultural Fields

Open Areas Water Bodies

Residential Areas Commercial Areas

Green Areas
Multispectral Image Interpretation Process
The Decision Tree for the general rule
base of the developed areas
3.5. Extracted Information
Multispectral Image Interpretation Results, 2000, Han. Landsat Data

Evaluation Technique
Polygon-Pixel Count 89.0%
Kappa Index 0.86

Reserved Areas Developed Areas

 Assignment 3
• Select one classifier algorithm from each category
namely per pixel, sub-pixel, per field and
contextual based and explain its operation and do
a comparison.
• Do a report on your findings using the template in
the next slide.

• Date of Issue: 07/10/2024

• Date of Submission: 28/10/2024
Item Per pixel Sub-pixel Parcel/Per field Contextual

Example

Function

Application

Advantages

Disadvantages
The End