Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing

3.

An introduction to Remote Sensing and Basic Principles of

GIS

Remote Sensing (RS) is the act of deriving data about the earths surface, from
above it. Previously images were obtained through photography from aircraft for
processing into maps. Now the term Remote Sensing is used most often specifically
to satellite-mounted multispectral1 sensors, such as the Landsat series of satellites.
This is normally available in digital format from one of the satellite agencies (see
chapter 5). Some kind of RS or GIS software is required to be able to read the image,
though once the basic map image has been produced, techniques other than GIS
can be used to process or manipulate it (see chapter 4).
Geographical Information Systems (GIS)2 are best described as software and
techniques for displaying and analysing information about the earth in a digital form.
GIS and RS software share many similar attributes and some software packages
claim to be both RS and GIS software. They are similar fields in that they both are
concerned with the digital representation of geographic phenomena and often both
employ the same spatial analytical techniques to manipulate the data.
Once the image (usually received as a set of layers) has been processed, there are a
number of ways in which it can be used and manipulated, using manual and/or digital
techniques. The research has established that a map useable for basic planning
purposes can be produced with just three layers of a basic Landsat image or tile
(approximately 1% of the data supplied).

Optical Remote Sensing

Remote sensing is simply the observation of an object from a distance. Various
forms of aerial photography have been used to create maps of the earths surface
since the eighteenth century. Satellite remote sensing can be traced back to the mid1940s and the launch of various rockets (the Viking series) from White Sands in New
Mexico, and which contained cameras on board. The first satellite sensor dedicated
to the multi-spectral remote sensing of earths surface was launched in 1972. Initially
named the Earth Resources Technology Satellites (ERTS), this series was renamed
Landsat in 1975, and is still in operation today.
Over the last three decades earth observation has become increasingly
sophisticated, with a large number of countries including those in the developing
world recently launching their own remote sensing satellites, such as India, Brazil,
China and Pakistan. Remote sensing data has proved invaluable in a wide range of
research fields, from making maps of the earth surface, to monitoring renewable and
1

Multispectral is the term used for recording images at various wavelengths, each as a separate layer
related to a given defined image boundary. Singly or in combination they reveal much information about
that portion of the earth's surface - what is on it, in it and what it is composed of. When combined and
compared with adjoining images a much broader picture can be constructed of the earth's surface.
2
A GIS is a sophisticated database management system designed to acquire, image, visualise, and
display spatially referenced (geographical) forms of data. John R Campbell - Interdisciplinary Research
and GIS - ASA Monograph 39, Routeledge 2002.
Max Lock Centre, University of Westminster
254575008.doc

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing
non-renewable natural resources, to urban planning. Its success is due to its unique
capability to provide near real-time information of the earths surface at a variety of
scales. In addition, the continuous nature of the data acquisition, the regular revisit
of the sensor to previously imaged areas of the earth, and the digital nature of the
data have all added to its success and reliability.

The Electromagnetic Spectrum

The basic principle underlying remote sensing is the measurement of
electromagnetic energy reflected or emitted by various features on the earths
surface or in its atmosphere. The entire array of electromagnetic radiation at its
various wavelengths is referred to as the electromagnetic spectrum. Wavelengths
are measured in micrometers (1 micrometer (m) = 10 6 meters).
Figure 3.1: Simple Diagram of the Electromagnetic Spectrum.

Source: http://imagers.gsfc.nasa.gov/ems/waves3.html

At one end of the spectrum are radio waves and microwaves that have very long
wavelengths. At the shorter end of the spectrum are x-rays and gamma rays, and in
the middle is the narrow range of optical wavelengths visible to human eyes. Various
sections of the spectrum are of use in remote sensing. These are primarily
atmospheric windows bands of wavelengths where the radiation passes easily
through the atmosphere to the surface of the earth. The most commonly used are
visible and infrared wavelengths:

Visible portion, consisting of blue, green and red: 0.4m to 0.7m

Infrared portion, consisting of near, short-wave, and thermal: 0.7m to 14m

A satellite sensing system will be designed to image a certain portion of the

spectrum, i.e. to record the amount of radiation reflected or emitted at various
wavelengths. In order to collect a sufficient amount of energy to provide a reliable
measurement, satellite sensors will record the reflectance over an interval of
wavelengths. These are referred to as channels or wavebands. The Advanced Very
High Resolution Radiometer (AVHRR) for example images at one visible channel
(0.53 0.73m), one near infrared channel (0.73 1.1m), two short-wave infrared
channels (1.58 1.64m and 3.5 3.9m), and two thermal infrared channels (10.3
11.3m and 11.5 12.5m). It is therefore a multispectral sensor, as it images over
more than one waveband of the electromagnetic spectrum.

Max Lock Centre, University of Westminster

254575008.doc

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing

Picture Elements or Pixels

A remote sensing image is made up of an array of square picture elements or pixels.
In an optical image the size of the ground area represented by a single pixel is
determined by the spatial resolution of the satellite. The spatial resolution can be
described simply as the smallest object that can be distinguished from its
surroundings. Thus for images produced by the AVHRR sensor, the pixel size is 1km
by 1km. In comparison the Enhanced Thematic Mapper (ETM) onboard Landsat 7
has a resolution of 15m. A pixel in this case thus represents an area of 15m by 15m
(or 225m2).

The Principles of GIS

A GIS (Geographical Information System) can be described as a computer mapping
and data storage system, in which data is tied to points, lines, polygons or pixels on a
computer display. Rather than just a 'dumb map', in which colours and symbols
represent geographical features, the user can interact to varying levels with a GIS.
For example, click the mouse on a line and discover the name of the road it
represents, its length, the last time it was surfaced, who carried out the surfacing
work, when it will need surfacing again, statistics about the number of road traffic
offences committed there in the last year. The possibilities are almost endless.
Additionally, a GIS can use the spatial information it contains in combination with
attribute data to perform spatial analysis. For example we may want to know, how
many Korean restaurants there are within a twenty-mile radius of a hotel, or how far
you could get in a car along a certain route. You could ask what could I see if looked
in this direction, stood at this point or where could you see me from?

Data Structures
The way a GIS does this is to use data structures to represent geographical object. It
does so in two main ways called 'raster' and 'vector'.
Raster Data Structure
In a raster data structure, reality is divided into an array of square pixels each of
which has a value. For example, 1 may be road, 2 pasture, 3 water, etc.
The image on the left below shows a raster depiction of a house, by a river with a
road serving the house and continuing on past. Each pixel has a code that
represents its land type. The advantage of this is that another layer can be placed
showing property value, for example, over this map and a calculation can then be
made pixel to pixel through the layers. This can be done with as many layers as
needed.

Max Lock Centre, University of Westminster

254575008.doc

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing
Figure 3.2: Raster Versus Vector Data Structure

Raster Data Structure

Vector Data Structure

Vector Data Structure

The other (and most common) form of GIS data structure is the vector data structure
or object approach. Geographical objects are constructed of point, lines and
polygons. In the view above the house is a polygon, constructed of lines that change
direction at points.
Figure 3.3: Vector Depiction of an Imaginary Area

In figure 3.3, an imaginary vector depiction shows the three components used
together to depict geographic objects. The points could represent cities, or oil wells
any discreet location or area that can be represented as a single point on a map,
depending on the scale of the map.
The line could represent a road, a river, the route of an exploration or advance any
linear feature. The line is constructed of a series of points, known as vertices. Each
vertex defines the point at which the line changes direction. In a GIS some
Max Lock Centre, University of Westminster
254575008.doc

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing
knowledge is built into these lines, so for example the GIS can tell which direction
along each line is 'forward' or 'backwards', known as the 'to' and 'from' nodes, or
which county, zone is on each side of the line.
The polygonal theme which could be counties, property parcels, electoral zone etcis constructed of lines (in turn defined by points) that join up to create a closed area.

Interoperability and Availability

Although most GIS packages use a combination of raster and vector data structure,
they often call the same things by different names, or, even worse, use the same
name to describe different things. Worse than that, GIS software manufacturers often
produce GIS outputs in forms that are not compatible with other software types.
This is supposed to have stopped happening. In 1994 the Open GIS Consortium
(OGC) was formed with a view to standardising (or at least making interoperable) all
geospatial data. Great strides have been made to achieve this aim by the OGC and
advances have been made by the OGC. However, although steps were made toward
interoperability, it still isn't a reality. Data still cannot be opened in forms from one GIS
in another. For example, a MapInfo tab file cannot be opened in an ESRI product and
vice versa; an ESRI shp file cannot be opened in MapInfo.
The problem is that it does not seem to be seen to as truly in the GIS industrys
interest (or nature) to be open or interoperable. As the GIS market continues to grow,
competition continues and GIS companies individually think they need the 'edge' of
some kind to ensure their survival or success.

Uncertainty and Metadata

The problem with the digital representation of reality in this way is that there will
always be some kind of abstraction. Reality is immensely complex and maps are
generally very simple models of it. A whole sub-discipline of GIS has arisen to
discuss the nature of and potential solutions to the uncertainty that arises from this
fact and the uncertainty associated with the data upon which GIS maps are based.
The problem is that uncertainty is the only really certain thing in cartography / GIS.
And the only true strategy for dealing with it is to acknowledge its existence and
make some attempt to quantify the level of error or uncertainty in an image.
The best way of doing this is through metadata roughly speaking, data about data.3
The idea is that when you create a map or an annotated image, you create a file that
goes with it the metadata file which explains, for example, where the image / map
came from, what was done to it and when. This allows anyone who uses the image
after you to quantify the error or uncertainty associated with the image. Figure 3.4
shows a metadata file for a Landsat image from ESDI.

metadata n., a set of data that describes and gives information about other data. OED

Max Lock Centre, University of Westminster

254575008.doc

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing
Figure 3.4: Metadata File for Landsat Image

However, some metadata standards can be extremely exhaustive. For example, the
document shown in figure 3.5 shows a file created to describe a GIS layer of wells
created by the University of Wyomings Water resources Centre
(http://www.wygisc.uwyo.edu/metadata/wells.html).
Figure 3.5: Metadata File for Well Head GIS Layer (University of Wyoming).

Classifying the Data

A primary use of remote sensing data is in classifying various features within a scene
into classes, thereby creating a thematic map. In a certain waveband or channel a
road, for example, will reflect less energy than the surrounding vegetation, whereas
in other wavebands it will reflect or emit more. Therefore, theoretically, it should be
possible to distinguish between different types of surface features based on their
spectral responses or spectral signatures.
The more wavebands that are used, the easier it should be to separate different
features, as the more detailed their spectral signatures will be. This is the theoretical
basis underlying multispectral classification procedures. Although a classification
may be performed on a multitemporal (as opposed to multispectral) dataset, this is

Max Lock Centre, University of Westminster

254575008.doc

10

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing
only of relevance if distinct spectral changes are expected to take place over time, for
example in agricultural seasons and crop rotation or rapidly developing or changing
urban areas.
The theory of classification allows us to perform two types of classification
unsupervised or supervised. The main difference between the two is that, during an
unsupervised classification, it is the computer that develops the spectral signatures
that will be used in the classification process, not the user. In addition, with an
unsupervised classification, pixels are assigned to a class based on their spectral
characteristics alone. During the unsupervised classification, pixels will be clustered
based on the natural spectral groupings present in the dataset.
Exactly how the assignment of a pixel to a cluster takes place will depend on the
software used. The most commonly used is the Iterative Self-Organizing Data
Analysis Technique (ISODATA) algorithm. Pixels are assigned to classes based on
their spectral distance from a class mean. This is an iterative process, with the class
means shifting after each iteration. The process ends when either a maximum
number of iterations have been performed, or a maximum percentage of unchanged
pixels has been reached between iterations. The various clusters created will then
need to be compared to in situ data, in order to assign meaningful values to them.
In contrast, a supervised classification requires some prior knowledge of the classes
present within the scene, in order to create a set of training data from which the
spectral characteristics of each class will be determined. The first stage is thus the
training stage. This involves the identification of areas representative of each class
within the image. The DN values of each area are then used to create spectral
signatures for each cover type. As the classification will be based on these training
sets, it is imperative that they represent the full spectral variation of each cover class.
The accuracy of the final classification is determined by the quality of these training
sets, so it is important to spend some time creating these as accurately as possible.
Once the training sets have been identified, the image can be classified. Numerous
supervised classification algorithms exist. Some of the most common ones that are
typically found in image processing software are described briefly in Boxes 3.1 and
3.2.

Max Lock Centre, University of Westminster

254575008.doc

11

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing

Box 3.1 Types of Classifier

The Minimum-distance-to-means classifier is one of the simplest classification
algorithms. The mean spectral value for each waveband, for each cover class
is computed. The distance between the value of an unknown pixel is then
compared to the mean values for each cover class. Finally, the pixel is
assigned to the class to which it is closest. If the pixel value is further from all
of the mean values than a user-defined distance, it is classified as unknown.
In contrast, the Parallelepiped classifier examines the range of the values in
each of the training sets. A parallelepiped (a rectangular area) is defined
surrounding the mean value of each class if the pixel value lies within the
parallelepiped it is assigned to that class. If the pixel falls within more than one
class, it is assigned to an overlap class, and if it falls in none, it is assigned to
unknown.
Finally, the Maximum likelihood classifier examines both the variance and
covariance of the spectral signatures for each cover class in order to classify a
pixel. The probability that a pixel belongs to a certain class is calculated, with
the pixel being assigned to the most likely class or labelled unknown if all the
probability values are below a user-defined threshold.

Max Lock Centre, University of Westminster

254575008.doc

12

Mapping Urbanisation for Urban and Regional Governance

Main Report
Final Report: September 2003 - DFID Research R8130
3. An Introduction to Remote Sensing

Box 3.2 Automatic Extraction of Features from Satellite Imagery is not always
Practicable
In remote sensing technology, the use of algorithms to automatically extract features from satellite
imagery in order to create maps, is still reasonably new and cutting edge. Armin Gruen in
Switzerland and Ian Dowman in the UK have worked extensively with algorithms that can follow
linear features or extract features based on their shape and / or spectral characteristics. This
represents a great advantage for the cartographic industry and national mapping agencies (NMAs)
have been very interested in their development. The drawback is that it is extremely expensive. The
effort required in terms of images, processing, hardware, training make this a very specialised
operation and so not really appropriate at present for the developing world. However, there is a
reference section at the end of the report giving the key references on automatic feature extraction
for mapping.
In practice, Landsat imagery, for instance, either printed out or viewed on a screen can be generally
and well interpreted at 1:25,000 to 1:50,000 scale by an operative trained in aerial photography and
who has a local ground knowledge of the area covered. Areas of uncertain classification such as
water, rock, building, road, railway, new grown grass, recently tilled land, bare earth, and so forth will
still need to be identified as a query and checked on the ground. The process can be iterative and
self correcting if clear records are kept.
There is a need to keep field checking of general topographical 'map' classification separate from
the collection of detailed 'quality' and 'quantity' land use classification. The latter should be a
separate exercise once the topographical base has been verified.

The advantages and disadvantages of the RS approach

Advantages

Disadvantages

It is relatively Inexpensive. The cost of software

and data (which often represents a one-off cost) is
less expensive than sending teams of surveyors
out into the field.
Current (within reason). One particular problem
that the developing world faces is that data is old or
out of date. Satellite imagery can be acquired for
free from the last decade and contemporary data
can be acquired fairly inexpensively
Provides data about large areas

There will be doubtful and uncertainties of

classification related to pixel size. A full field check
will be necessary to resolve ground use in these
areas.
Datasets from multiple sources are sometimes
difficult to georeference. Using images and maps
that are drawn in different scales and projections
can lead to difficulty combining them.

Provides data about inaccessible areas - or even if

theyre not inaccessible, then at least you dont
have to go there.
Rapid production of maps possible
Easy to manipulate (relatively!) with computers and
derive information for map production.

The sensor performs a sweep and as such can

create errors. i.e., only some of the image is
directly below the sensor and so pixels toward the
edge of the image may be distorted.
Objects in the image can be confused or misclassified. For example, shadows may look like
metalled roads.
To get any level of detailed, current data and to buy
specialist RS software can be expensive
In some active sensing systems (such as lidar) the
sensor and source are moving relative to each
other distortions can creep to the image. This is a
form of Doppler effect (the visual equivalent of a
police siren changing pitch as it moves closer or
further away from you).

Rapid collection of data much more efficient that

ground survey!

Max Lock Centre, University of Westminster

254575008.doc

13