remote sensing-based spatial information - food … management of spatial information unit3: spatial...

Management of Spatial Information

Remote Sensing-Based Spatial

Information

Learners’ Notes

© FAO - ITC 2013

Course: Management of Spatial Information Unit3: Spatial Analysis

Lesson 1: Remote Sensing-Based Spatial Information

Learners’ Notes 1

Table of contents Learning objectives ................................................................................................. 2

Introduction ........................................................................................................... 2

Information extraction methods .............................................................................. 3

Image interpretation ............................................................................................... 3

Image interpretation – basic elements ..................................................................... 4

Image interpretation – the process .......................................................................... 6

Fieldwork – sample design ...................................................................................... 7

Digital image classification ...................................................................................... 8

Digital image classification – feature space .............................................................. 8

The overall objective of cluster definition ............................................................... 10

Digital image classification - algorithms ................................................................. 10

Digital image classification - validation of the result ................................................ 12

Digital image classification - pixel-based ................................................................ 13

Digital image classification - object oriented ........................................................... 15

Sources to obtain remotely sensed data ................................................................. 16

Remote sensing products ...................................................................................... 17

Summary ............................................................................................................. 18



Learners’ Notes 2

Learning objectives

At the end of this lesson, you will be able to:

• define the general mapping methodology with remotely sensed data for land cover;

• describe the visual image interpretation process and key elements;

• describe the field data collection methods;

• describe digital image classification activities including validation of results; and

• find some sources for remote sensed data and products.

Introduction

Spatial information as derived from remote sensing (RS) uses a general mapping methodology

that consists of:

• the interpretation of images for sampling design;

• field data collection; and

• analysis and information extraction.

The assumption in this mapping methodology is that areas that look homogeneous in the image

have similar features on the ground. Maps and inventories should reflect what is actually on the

ground (the ‘ground truth’).

Therefore, field visits should be made to observe what is there in reality.

Field visits for ground observation are time-consuming and usually costly. Making observations

everywhere in the entire area to be mapped is likely to take too much time.

For efficiency reasons, remote sensing data is used to extrapolate the results of a limited number

of observations over the entire area at study.

Based on the interpretation or classification, a field sampling design is made.

Once the correlation between the collected field data and the interpretation/classification is

established and understood, the entire area can be mapped in terms of what is on the ground.

The actions to be taken depend on the quality of the correlation:

• if it is of good, only recoding and renaming of the units is required; while

• if it is poor, a complete re-interpretation or classification might be needed.



Learners’ Notes 3

Information extraction methods

In general, the information extraction methods from RS images form two groups:

1. information extraction based on visual image interpretation (e.g. for land use or soil

mapping and acquisition of data for topographic mapping from aerial photographs); and

2. classification based on semi-automatic processing by the computer (e.g. digital image

classification and calculation of surface parameters).

Image interpretation

The most intuitive way to extract information from remote sensing images is by visual image

inspection and interpretation, which is based on the human ability to relate colours and

patterns in an image and to real world features.

We can interpret images displayed on a computer monitor or images given in a hardcopy

form.

Delineation can be done:

• on a transparency overlaid on an aerial photograph or hardcopy of the satellite image; or

• by digitizing either on-screen, or using a digitizer tablet when we have a hardcopy image.

Instead of interpreting and digitizing on a single image, we can use a stereo-image pair. The

interpretation process is the same, we only need special devices for stereoscopic display and

viewing, and equipment that allows us to properly measure in a stereogram.

Visual image interpretation is not as easy as it may seem; it requires training. Yet the human eye-

brain system is well capable of doing the job. In analysing an image, typically you can be

somewhere between the following two situations:

Direct and spontaneous recognition - Direct and spontaneous recognition, refers to the ability

of an interpreter to identify objects or features at first glance. Agronomists will immediately

recognize the pivot irrigation systems with their circular shape. They are able to do so because of

earlier (professional) experience.

Similarly, most people can directly relate what they see on an aerial photo to the terrain features of

their place of living (because of ‘scene knowledge’).



Learners’ Notes 4

The quote from people that are shown an aerial photograph for the first time “I see because I

know” refers to this spontaneous recognition.

Logical inference - Logical inference, needing several clues to draw conclusions through a

reasoning process. Logical inference means that the interpreter applies reasoning.

In the reasoning, the interpreter uses acquired professional knowledge and experience.

What is the white rectangle?

Based on the connecting white line (probably track) from another straight line (probably road) the

conclusion can be that it is a house or a building.

Image interpretation – basic elements

You can distinguish the following basic visual interpretation elements:

Grey tone/

colour

The tone is directly related to the amount of light reflected from the

surface.

Variations in colour are primarily related to the spectral characteristics of

the terrain features and the bands selected for visualization.

Texture It relates to the frequency of tonal change. Texture may be described by

terms as:

• coarse or fine;

• smooth or rough;

• even or uneven;

• mottled;

• speckled;

• granular;

• woolly;

• etc.

Pattern

It refers to the spatial arrangement of objects and implies the

characteristic repetition of certain forms or relationships. Pattern of

different land uses in Java Indonesia. A rice fields, B salt/fish ponds and C

villages.

Pattern can be described by terms such as:

• concentric;



Learners’ Notes 5

• drainage;

• dot;

• field;

• etc.

Shape Shape or form characterizes many objects visible in the image.

Shape and size are mainly applicable for man-made features.

Size Size of objects is a function of scale. It can be considered in a relative or

absolute sense.

Height Height differences are important for distinguishing between different vegetation types, building types, etc.

Height can be used if we have stereo images or photographs.

For stereovision you need special techniques.

Shadow of objects can also indicate height.

Location/association Location/association refers to the situation in the terrain or in relation to

its surrounding. A forest in the mountains is different from a forest close

to the sea or near the river in the lowland.

Looking at the main seven interpretation elements you may have noticed a relation with the

spatial extent of the feature to which they relate:

• tone/colour can be defined for a single pixel1

• texture is defined for a group of adjacent pixels, not for a single pixel;

;

• the other elements relate to individual objects or a combination of objects.

The simultaneous and often implicit use of all these elements is the strength of visual image

interpretation.

In standard digital image classification, only colour is utilized, which explains the limitations of

automated methods compared to visual image interpretation.

1 Pixel - Contraction for picture element, which is the elementary unit of image data. The ground pixel size of

image data is related to the spatial resolution of a sensor system it was produced by.



Learners’ Notes 6

Image interpretation – the process

The interpretation process consists of delineating areas that internally appear similar to, yet, at

the same time different from other areas.

Making an interpretation of only one aerial photograph or a small part of an image from a

spaceborne sensor seems quite simple. You have the overview of the entire area at all times and

can easily compare one unit with another, and decide if they are the same or different.

However, working with many photographs and also with several people will require a good

definition of the units to be delineated.

The definition of units is based on what can be observed in the image.

Different interpretation units can be described according to the interpretation elements.

After establishing what the features are on the ground, ‘interpretation keys’ can be constructed.

On the basis of these keys, an interpretation of features can be made.

These features are again described in terms of interpretation elements , as showed in the images

beside.

After fieldwork, it should become clear, what the units represent on the ground.

Prior to the delineation of the units, a legend is constructed based on interpretation elements.

The legend can be presented in the form of a table, in which each element type is represented by a

column.

You can see in the table presented here, a theoretical example of such a legend description. In this

legend, the unit number represents a yet unknown feature type and the corresponding row

elements will be used to identify that feature type.

Unit Tone Texture Pattern Shape Size Height Location

1 black

2 grey smooth

3 grey rough high mountain

4 grey rough low

5 grey rough hihg sea + river

6 grey

white

Field line



Learners’ Notes 7

7 white

grey

field

line

8 grey+black field square

9 grey+black field rectangle 5 x 5

10 grey+black field rectangle 20 x20

Fieldwork – sample design

As you already know, field visits for ground observation are time-consuming and can be costly. The

selection of sample locations is a crucial step to make mapping cost-effective.

You can use the RS images to stratify the area, using a method of sampling called stratified

sampling. With this method you can make an interpretation of the area to be mapped based on

the interpretation elements. The interpretation units are the strata to be sampled. In all strata, an

equal amount of samples is taken.

You can also select the samples in such a way that they are representative for the interpretation

elements of their unit. This is called stratified representative sampling.

Stratified sampling is a very time-efficient and cost-effective method as compared to random

or systematic sampling. If an interpretation unit occupies only a very small area:

• with random or systematic sampling, many samples are needed to ensure that it is

sampled; while

• applying the stratified sampling approach, far fewer samples are needed.

Stratified representative sampling can only be applied if the data to be obtained is qualitative (i.e.,

nominal or ordinal). For mapping of quantitative data (i.e., interval or ratio data), unbiased

sampling strategies (i.e., random or systematic sampling) should be applied to allow statistical

analysis.

Example of quantitative data: biomass measurements

In biomass measurements the entire area needs to be sampled and no prior interpretation is

needed for the sampling strategy. Both the stratified and unbiased sampling strategies are used if

quantification of certain strata is not required. For instance, we use stratified random sampling of

grass biomass for livestock management when in the strata forest, water, and urban areas, no

biomass measurements are needed. We do so to limit the time-consuming unbiased sampling.



Learners’ Notes 8

During fieldwork, we need to know where we are in relation to the RS image, so we can relate

features in the field to objects in the image. This is to improve our understanding of the image

data, and because we use the image for stratification.

Mobile GIS helps tremendously in this, since we see our position directly in the image. After

establishing the location of the sample, we can collect the required field data.

Digital image classification

Image classification uses the differences in spectral characteristics of materials of the Earth's

surface.

Multispectral image data captures some of those differences. Here you will focus on the

classification of multispectral data.

The principle of image classification is that a pixel is assigned to a class on the basis of its

combination of spectral band values. Doing so for all pixels will result in a classified image.

The crucial point of image classification is in comparing the pixel values to predefined clusters,

which requires definition of the clusters and methods for comparison. Definition of clusters is

a process requiring user interaction.

Comparison of the individual pixels with the clusters takes place using classifier algorithms.

A digital image is a two-dimensional array of pixels. The digital number (DN) value of a pixel

lies in the range 0 to 255, when 8 bits are used for its encoding. A DN corresponds to the EM

radiation reflected from the pixel’s area.

Digital image classification – feature space

Feature space - The mathematical space describing the combinations of observations (DN values

in the different bands) of a multispectral or multi-band image. A single observation is defined by a

feature vector.

When we consider a two-band image, we can say that the two DN values for a pixel are

components of a two-dimensional vector [v1, v2], the feature vector.

An example of a feature vector is [13, 55], which tells the conjugate pixels of band1 and band2

have the DN’s 13 and 55. This vector can be plotted in a two-dimensional graph.



Learners’ Notes 9

Similarly, we can visualize a three-dimensional feature vector [v1, v2, v3] of a cell in a three-

band image in a three-dimensional graph.

A graph that shows the feature vectors for an image is called a feature space, or ‘feature space

plot’ or ‘scatter plot’.

Plotting the values is difficult for a four- or more-dimensional case, even though the

concept remains the same.

A 2D histogram provides information about pixel value pairs that occur within a two-band image.

Plotting all frequencies of feature vectors of digital image pair yields a 2D histogram of many points.

This image shows the frequencies of red and infrared bands of a SPOT image of Enschede (NL).

Vector plots in red and yellow are most frequent, or in other words, the combination of values is

found often in the image.

Purple combinations occur little and black combinations do not exist in the image.

The image below shows a feature space in which the feature vectors are plotted for samples of five

specific land cover classes (grass, water, trees, etc). Fairly obviously, the feature vectors of pixels

that are water areas form a compact cluster. Also, the feature vectors of the other land cover types

(classes) are clustered.

The image illustrates the basic assumption for image classification: a specific part of the feature

space corresponds to a single class.

Once all class boundaries have been defined in the feature space, each feature vector of a multi-

band image can be plotted and assigned to the class in which it fits best.



Learners’ Notes 10

The overall objective of cluster definition

The overall objective of cluster definition is to assemble a set of statistics that describe the spectral

response pattern for each land cover type that occurs in the image.

Clusters are defined by selecting homogeneous areas in the image and assigning a class name.

Classes should be well separated from each other.

Display of different sample classes in the feature space assists in the evaluation of the clusters.

We use distance or cluster parameters in the feature space to accomplish classification.

Distance in the feature space is measured as Euclidian distance, using the unit of the DN axes. In

a two-dimensional feature space, the distance can be calculated with Pythagoras' theorem.

Cluster parameters are minimum, maximum and mean.

Digital image classification - algorithms

After the clusters have been defined, classification of the image can be carried out by applying a

classification algorithm.

Several classification algorithms exist. The choice of algorithm depends on the purpose of the

classification and the characteristics of the image and training data.

There are three classifier algorithms:

1. Box classifier;

2. Minimum Distance to Mean (MDM); and

3. Maximum Likelihood (ML).

In practice, the box classifier is hardly used. The MDM and the ML classifiers are used most

frequently.

The Box classifier is the simplest classification method. With it, upper and lower limits are

defined for each band and class.

BOX CLASSIFIER

The limits may be based on:

• the minimum and maximum values; or




• the mean and standard deviation per class.

When the lower and upper limits are used, they define a box-like area in the feature space. The

number of boxes depends on the number of classes.

During classification, every feature vector of an input (two-band) image cell is checked for

containment in any of the boxes If so, the cell will be assigned the class label of that box. Cells that

do not fall inside any box are assigned the `unknown class'.

A disadvantage of the box classifier is the possible overlap between classes. When a cell is a

box overlap, it is arbitrarily assigned the label of any of those boxes.

The basis for the MDM classifier is the set of cluster centres.

MINIMUM DISTANCE TO MEAN (MDM)

During classification, the Euclidean distances from a candidate feature vector to all the cluster

centres are calculated. The candidate cell is assigned to the class with shortest distance.

Some disadvantages of the MDM classifier are the following:

• points that are at a large distance from a cluster centre can still be assigned to this centre

(this problem can be overcome by defining a threshold value that limits the search

distance); and

• it does not take class variability into account, some clusters are small and dense, while

others are large and dispersed.

The Maximum Likelihood (ML) classifier does not only consider the cluster center point, but also

the cluster shape, size and orientation. This is achieved by determining a statistical distance

and a covariance matrix for the cluster.

MAXIMUM LIKELIHOOD CLASSIFIER (ML)

The assumption of most ML classifiers is that the statistics of the clusters follow a normal

(Gaussian) distribution.

For each cluster, so-called equiprobability contours can be drawn around the centers of the

clusters.

Progressively larger ellipses surrounding the center represent contours of probability of membership

to a class, with the probability decreasing away from the center. Maximum likelihood also allows the

operator to define a threshold distance by defining a minimum probability value.




The statistical distance - It is a probability value, the probability that observation x belongs to

the cluster. A cell is assigned to the class (cluster) for which it has the highest probability.

Digital image classification - validation of the result

Image classification results in a raster file in which the individual raster elements are class

labelled.

As image classification is based on samples of the classes, the actual quality of the classification

result should be checked. This is usually done by a sampling approach in which a number of raster

elements of the output is selected and both the classification result and the true world class are

compared.

Comparison is done by creating a correlation matrix from which different accuracy measures can

be calculated. The `true world class' is preferably derived from field observations.

Sometimes, sources of an assumed higher accuracy, such as aerial photos, are used as a reference

instead of field observations.

Once the sampling for validation has been carried out and the data collected, a correlation

matrix (or error matrix) can be established. In the table, four classes (A, B, C, D) are listed. A

total of 163 samples has been collected.

A B C D Total

Error of Commission (%)

User Accuracy (%)

a 35 14 11 1 61 43 57

b 4 11 3 0 18 39 61

c 12 9 38 4 63 40 60

d 2 5 12 2 21 90 10

Total 53 39 64 7 163

Error of Omission

34 72 41 71

Producer Accuracy

66 28 59 29

From the table you can read that:

• 53 A cases were found in the real world (`reference'); while




• the classification result yields 61 cases of A; and

• in 35 cases, they agree.

The first and most commonly cited measure of mapping accuracy is the overall accuracy, or

Proportion Correctly Classified (PCC).

Overall accuracy

It is the number of correctly classified pixels (i.e., the sum of the diagonal cells in the error matrix)

divided by the total number of pixels checked. The overall accuracy yields one figure for the result

as a whole.

Most other measures derived from the error matrix are calculated per class.

Error of omission refers to those sample points that are omitted in the interpretation result.

Consider class A, for which 53 samples had been taken. Some 18 out of the 53 samples were

interpreted as b, c or d. This results in an error of omission of 18 / 53 = 34%. Error of omission

starts from the reference data, and therefore relates to the columns in the error matrix.

The error of commission starts from the interpretation result and refers to the rows in the error

matrix. The error of commission refers to incorrectly classified samples. Consider class d: only two

of the 21 samples (10%) are correctly labelled. Errors of commission and omission are also referred

to as type I and type II errors, respectively.

Another widely used measure of classification accuracy derived from the error matrix is the kappa

or k statistic.

Kappa statistics take into account the fact that even assigning labels at random results in a certain

degree of accuracy.

Based on Kappa statistics, one can test if two data sets, e.g. classification results, have different

accuracy. This type of testing is used to evaluate different RS data or methods.

Digital image classification - pixel-based

Pixel-based image classification is a powerful technique to derive thematic classes from multi-

band images. However, it has certain limitations that you should be aware of.

The most important constraints are that:

1. it results in spectral classes; and




2. each pixel is assigned to one class only.

Spectral classes are classes that are directly linked to the spectral bands used in the classification.

In turn, these are linked to surface characteristics. In this respect, you can say that spectral

classes correspond to land cover classes.

In the classification process, a spectral class may be represented by several training classes.

Among others, this is due to the variability within a spectral class.

Consider a class such as ‘grass’; there are different types of grass, which have different spectral

characteristics (see sample areas 7 and 21 in the SPOT image). Furthermore, the same type of

grass may have different spectral characteristics when considered over larger areas due to, for

example, different soil and climate conditions.

Sometimes one could be interested in land use classes rather than land cover classes.

Sometimes, a land use class may comprise several land cover classes. The following table here

gives some examples of linking spectral land cover and land use classes.

Between two columns there can be 1-to-1, 1-to-n, and n-to-1 relationships. The 1-to-n

relationships are a serious problem and can only be solved by adding data and/or knowledge to the

classification procedure.

The data added can be:

• other RS images (other bands, other moments) or

• existing geospatial data, such as topographic maps, historical land inventories, road

maps, and so forth.

Usually, this is done in combination with adding expert knowledge to the process.

Examples

An example is using historical land cover data and defining the probability of certain land cover

changes. Another example is to use elevation, slope and aspect information. This will prove

especially useful in mountainous regions where elevation differences play an important role in

variations in surface cover types.

The other main problem and limitation of pixel-based image classification is that each pixel is

only assigned to one class. This is not a problem when dealing with (relatively) small pixels.

However, when dealing with (relatively) large pixels, more land cover classes are likely to occur




within a cell. As a result, the value of the pixel is an average of the reflectance of the land cover

present within the pixel.

In a standard classification, these contributions cannot be traced back and the pixel will be assigned

to one of either classes or even to another class. This phenomenon is usually referred to as the

mixed pixel or mixel.

This problem of mixed pixels is inherent to image classification: assigning the pixel to one thematic

class.

The solution to this is to use a different approach, for example, assigning the pixel to more than

one class.

The problem of mixed pixels also highlights the importance of using data with the

appropriate spatial resolution.

As mentioned earlier, the choice of classification approach depends on the data available, but

also on the knowledge we have about the area under investigation:

• without knowledge of the present land cover classes, unsupervised classification (i.e.

classification without user interaction) can give an overview of the variety of classes in an

image;

• if knowledge is available (such as from field work or other sources) supervised

classification may be superior.

However, both methods only make use of spectral information, which gets increasingly problematic

with higher spatial resolution.

For example, a building constructed from different materials leads to pixels with highly variable

spectral characteristics, and thus a situation in which training pixels are of little help. Similarly, a

field may contain healthy vegetation pixels as well as some of bare soil.

Digital image classification - object oriented

We are also increasingly interested in land use. However, to distinguish for example urban from

rural woodland, or a swimming pool from a natural pond, an approach similar to the visual

interpretation is needed.

Object-oriented image analysis (OOA), also called segmentation-based analysis, allows us to

do that. Instead of trying to classify every pixel separately and only based on their spectral




information, OOA breaks down an image into spectrally homogeneous segments that

correspond to fields, tree stands, buildings, etc. It is also possible to use auxiliary GIS layers, for

example building footprints, to guide this segmentation.

With visual image interpretation we consider each element in terms of its spectral appearance but

also in terms of its shape and texture, and within its environment.

Similar to the cognitive approach of visual image interpretation in OOA we can specify contextual

relationships and more complex segment characteristics to classify the objects extracted in the

segmentation process.

What OOA is particularly suitable for?

OOA is particularly suitable for:

• images of high spatial resolution; and

• data obtained by airborne laser scanner or microwave radar.

It requires that we have substantial knowledge on what distinguishes a given land cover or land use

type, as well as auxiliary data such as elevation, soil type or vector layers

Sources to obtain remotely sensed data

The following are some sources for obtaining remotely sensed data:

• The USGS Global Visualization Viewer is a quick and easy online search and order tool

for selected satellite (Landsat, MODIS, ASTER) and aerial data (http://glovis.usgs.gov/).

• Spot Image is a leading Earth observation services company providing imagery products

and solutions to customers worldwide since 1986 (http://www.spotimage.com). Images can

be browsed at: http://sirius.spotimage.fr/PageSearch.aspx

• VITO processes, archives and disseminates Global 10 days, 1 KM S10/D10 products

derived from the SPOT 4/5 VEGETATION instruments (http://free.vgt.vito.be ).

• The CGIAR-CSI GeoPortal provides SRTM 90m Digital Elevation Data for the entire world

(http://srtm.csi.cgiar.org ).

• DigitalGlobe is an imagery provider for high resolution images and geospatial data

captured by the satellites Worldview and QuickBird (http://www.digitalglobe.com )

http://glovis.usgs.gov/�

http://www.spotimage.com/�

http://sirius.spotimage.fr/PageSearch.aspx�

http://free.vgt.vito.be/�

http://srtm.csi.cgiar.org/�

http://www.digitalglobe.com/�




• GeoEye is a provider of satellite and aerial imagery, location information products and

image processing services captured by the satellitse Geoeye and Ikonos

(http://www.geoeye.com).

Remote sensing products

The Group on Earth Observations (GEO) / Integrated Global Observing Strategy (IGOL)

Agricultural Monitoring Community of Practice was established in July of 2007 at the second

IGOL/GEO workshop convened at the headquarters of the UN Food and Agriculture Organization

(FAO) in Rome.

Agricultural Monitoring

This community of practice represents 25 national and international organizations concerned with

agricultural monitoring.

Its purpose is to develop and implement a strategy for global agricultural monitoring in the

framework of GEO (http://www.earthobservations.org/cop_ag_gams.shtml ).

The Global Land Cover Facility (GLCF

Land cover

http://glcf.umiacs.umd.edu/index.shtml) provides earth

science data and products to help everyone to better understand global environmental systems.

In particular, the GLCF develops and distributes remotely sensed data and products that explain

land cover from the local to global scales. Primary data and products available at the GLCF are free

to anyone via FTP.

The FAO Remote Sensing Survey portal. As part of the Global Forest Resources Assessment

2010 (FRA2010), FAO and its member countries and partners undertook a global remote sensing

survey of forests.

Forest Monitoring

The FRA 2010 Remote Sensing survey builds on the experiences from the remote sensing surveys

of the tropical region part of previous global forest resources assessments and on recent advances

in methodologies and availability of imagery.

FAO and its partner organizations have made pre-processed imagery easily available through the

internet access to free remote sensing data and specialized software, this will particularly benefit

http://www.geoeye.com/�

http://www.earthobservations.org/cop_ag_gams.shtml�

http://glcf.umiacs.umd.edu/index.shtml�




developing countries with limited forest monitoring data or capacity.

http://geonetwork4.fao.org/geonetwork/srv/en/fra.home

The Global Fire Monitoring Center (GFMC

FireMonitoring

www.fire.uni-freiburg.de/). Following the

recommendations of the UN-ECE/FAO/ILO Seminar Forest, Fire and Global Change (Russia 1996)

and a number of international conferences, the UN- ECE/FAO Team of Specialists on Forest Fire

proposed the establishment of an institution which at that time was preliminarily designated as a

Global Fire Management Facility.

On the basis of these recommendations the Government of Germany through the Ministry of

Foreign Affairs, Office for the Coordination of Humanitarian Assistance, in June 1998 provided

initial funding for the establishment of such an entity which was designated Global Fire Monitoring

Center (GFMC).

The GFMC was inaugurated at the FAO Meeting on Public Policies Affecting Forest Fires (Rome,

October 1998).

Summary

The general mapping methodology with remotely sensed data for land cover involves: (a)

interpretation of images for sampling design; (b) field data collection; and (c) analysis and

information extraction.

Information extraction can be based on visual image interpretation and classification based on

semi-automatic processing by the computer.

Image interpretation requires training and needs a systematic approach. There is a set of

interpretation key elements that express the characteristics of an image.

In filed data collection, the selection of sample locations is a crucial step to make mapping cost-

effective.

The principle of image classification is that a pixel is assigned to a class on the basis of its

combination of spectral band values. Doing so for all pixels will result in a classified image.

Comparison of the individual pixels with the clusters takes place using classifier algorithms.

Image classification results in a raster file in which the individual raster elements are class

labelled.

http://geonetwork4.fao.org/geonetwork/srv/en/fra.home�

http://www.fire.uni-freiburg.de/�




Pixel-based image classification is a powerful technique to derive thematic classes from multi-

band images.

Object-oriented image analysis (OOA) breaks down an image into spectrally homogeneous

segments that correspond to Earth’s elements.

remote sensing-based spatial information - food … management of spatial information unit3: spatial...

Documents