A landscape classification can be produced by combining the four main attributes of landscape discussed in the previous chapters. The unique combination of these attributes at any chosen level of generalisation forms the basis of individual landscape classes. Figure 6.1 shows graphically this combination process for generalisation level 3. Here, the resulting landscape classification has a total of 536 unique classes, and a total of 3115 discrete areas. It is not feasible to produce a key for this many classes, and therefore Figure 6.2 shows a key for only the top ten classes in total area. A classification code, L3 V3 N3 W3, shows the generalisations used. It means that generalisation level 3 was used for all four attributes: landform (L), vegetation (V), naturalness (N), and water (W).
It should be noted that the results near
the boundary of the study area are inaccurate because the classification
uses neighbourhood information. Near the boundary, the information beyond
the boundary is not available, therefore the classification applied here
is inconsistent compared to the centre of the study area. The extent of
this inaccuracy is 10km (marked in Figure
6.2 by the inner square), since this was the extent of the largest
focal radius used.
6.2 Generalisation
Various levels of generalisation can be obtained by combining different generalisation levels of the landscape attributes. Figure 6.3 shows graphically six different outcomes, and the number of discrete areas and unique classes that have resulted from these. The most detailed level has 1302 classes, and the most general level has only 56 classes. It is possible to produce different outcomes by combining various levels of different generalisations. For example, if landform is considered more important than the rest, then level one of landform could be combined with less detailed levels of the other attributes, such as, a L1 V5 N5 W5, or a L1 V6 N2 W6 combination. Thus a range of classifications can be obtained that reflect different generalisations.
The number of classes identified in Figure 6.3 is only the number identified in the study area. The classification has the potential to identify many more. For level one, there is the potential for 356,224 classes to be identified (the product of 22 landform classes, 46 vegetation classes, 22 naturalness classes, and 16 water classes). However, this is only the tip of the iceberg, because considerable generalisation was required even to produce level one. Without this generalisation the classification would have the potential to produce approximately 6.7 X 1017 different classes.
The question then becomes: What level of generalisation
is appropriate? This depends partly on the scale at which the classification will be used
(ie. international, national, regional, or local), and partly on the purpose of the
classification at the chosen scale. Selection of an appropriate level for a particular
investigation might require preliminary cognitive and psychophysical research. The
variability in such research between areas of the same class, will demonstrate whether the
classifications are actually distinguishing the necessary subtleties. If two areas of the
same class are perceived as being different in terms of quality, then the classification
has not distinguished the necessary subtleties required. It is difficult to ascertain what
subtleties are important at the different levels of investigation.
One can speculate on the appropriate generalisation for different levels of scale from two different approaches. It can be said that there is a certain limit to the number of classes that research can cope with, especially for doing psychophysical preference surveys. When surveying public preferences using photos, it is practically not feasible to ask people to rank more than thirty photos (Auckland Regional Authority, 1982). The preferences of different groups of photos can be linked together by having some photos that are common to each group. Therefore, many groups of photos can be used. However, there would be a practical limit to this. The other approach that can be used to determine the appropriate number of classes is to decide what landscape components are really essential in a landscape classification and then to include only these. Large components like mountains, hills and plains, coast, lakes, urban areas, and areas of forests, and grassland have a significant visual impact and therefore should be included. Also, distinctions based on naturalness should be kept since this is a known contentious characteristic. However, because of the lack of landscape content category research in New Zealand, it is difficult to reason with some substantive evidence about this. Trial and error (hypothesis testing) is the only scientific method for determining the appropriate level of generalisation.
Once the landscape attributes have been combined, there
is further opportunity for generalisation using definitions based on more than one
attribute. Some classes of one attribute may be considered unimportant when a class of
another attribute is present. For example, in mountainous regions, it may be considered
unnecessary to include rivers since the two are often associated with each other, and
perhaps, because mountains are so dominating visually, rivers become insignificant. Zube
has suggested the following:
"As landform increases in dimension from flatlands
through hills to mountains, land pattern decreases in importance as an element of visual
quality. And, as landform decreases, the diversity of land pattern becomes increasingly
important as an element of visual quality." (Zube, 1984b, p. 122) With the combined coverage, such associations can be
identified and generalised. For example, two landscape classifications can be developed
with one having detailed vegetation information and the other not. A third landscape
classification can be produced by mixing these two classifications so that general
vegetation classes exist where there are mountains, and detailed vegetation classes exist
where there are not. This has not been done but the option is there. It is possible to do
this kind of generalisation before combination by doing a query on two different attribute
coverages, for instance "if mountains are present in the landform coverage and river
is present in the water coverage then change the water coverage". Yet another way of generalising is to use a neighbourhood
majority function. This replaces the central cell with the class that has the majority of
area within a defined search radius. This has the effect of removing the smaller discrete
areas. Just because a landscape class occupies only a small area does not mean that it is
unimportant. Small size might contribute to its significance, especially if it is unique.
It may not therefore be appropriate to generalise using such a filter. It seems appropriate to generalise the individual
landscape attributes before combining them to create a landscape classification. At that
stage of the process, the number of classes is more manageable, and the problem is divided
into smaller problems. Trying to develop a process that generalises a coverage that has
the potential to have 6.7 X 1017 classes would be impractical. 6.3 The application of an agreement model In section 5.4.2,
a method for incorporating fuzzy set theory was demonstrated on a landform
classification using entropy and agreement models. This appears useful
for researchers for ascertaining the degree of certainty of the classes
identified for different areas. Researchers would then be able to locate
study areas where there is a high (or low if this is appropriate) consensus
over their identification. It was concluded that agreement models are
the preferred approach for landform classification. An agreement model
could also be applied to the total landscape classification that has been
developed. An estimate of the processing time required to do this for
the study area would be approximately one week using a Sun Sparc 10 with
three parallel processors. This would be if 45 different classifications
were produced for each of the four different attributes, which, when combined,
produce over four million landscape classification, each with six different
levels of generalisation. The overall agreement model could be produced
by multiplying together the proportion values of the agreement models
of the four individual attributes, and therefore only 180 attribute classifications
would need to be produced, rather than four million landscape classifications.
This has not been demonstrated in this study because of the amount of
processing required. 6.4 Validity Now that a landscape classification
process has been developed and applied to the study area, it is time to
discuss the validity of this process and the resulting classifications.
In sections 2.8 and 2.9 two sets of criteria were
established for this purpose. They are based on general classification
principles, and specific landscape criteria that consider the important
characteristics of landscapes. In section 3.5.2,
it was argued that using these criteria, along with consideration of GIS
errors, was the most appropriate means available for this study for assessing
the validity of the landscape classification process. A comparison between
automated and manual classification, based on these criteria, will indicate
whether there has been an improvement. 6.4.1 General classification criteria Is the classification exhaustive and mutually
exclusive? The classification is exhaustive for the study area.
However, without modification it would not be exhaustive for all areas, especially areas
outside New Zealand. The classification has been designed for the study area. In some
other areas there will be different compositions of landscape components that have not
been catered for and modification would therefore be needed. This is particularly the case
with vegetation, which has many different components. The classification is mutually
exclusive for the study area. As described in section 4.2.3.
checks were made to ensure that an area could not be defined to more than
one class. Is the classification easily understood and applied?
It is questionable whether this classification is easily
understood by researchers who have had no training in the concepts of GIS. To them, this
classification may appear very complex. The actual fundamentals of this classification are
not complex. Basically, it is centred around the use of a focal neighbourhood mean
function, which in principle should be easily understood. This is applied to the
components of landscapes to identify compositions. The classification is complicated by
the lack of existing coverages for two of the landscape components: macro landforms, and
indented coasts. Considerable processing has been required to create these. Also, most of
the landscape component classes that were available needed considerable generalisation. The programs written for this classification have been
designed for research in order to enable maximum flexibility to explore different options.
Once a classification process has been developed, and agreed upon, there are two matters
that could be developed to improve the user friendliness of it. Firstly, a user friendly
interface can be developed that makes the classification easy to use. Secondly, the type
of NDDBs that the process is dependent on need to be standardised. If the topographical
databases used standard labels and identified standard entities, then this would make it
easier to develop a user friendly classification. Standardisation is now a major
consideration of cartography and will most likely be widely applied (Buttenfield and
McMaster, 1991). Once a user friendly classification has been developed, then a user would
state the names of the input files and the automated classification would do the rest. As for resources, the automated approach is very quick,
therefore requiring minimum human input. The whole of New Zealand could now be classified
within a few days. However, the classification does require the uses of expensive computer
resources - both hardware and software, and expensive databases. It can be argued that
computer resources are becoming cheaper all the time. Also, many resource management
institutions already have the necessary computer and data resources, but they are not
using them to full capacity. This classification can be run in batch mode in off peak
periods. The most human intensive and computer intensive part of automation is usually
developing the classification, not applying it. Now that a classification has been
developed, it can be easily applied. Is the classification repeatable and independent of
the researcher? Objectivity is one of the main advantages of automation.
The classification process only requires the researcher to start it. The rest is done by
the computer. It does not matter who starts the classification as the result will be the
same. The design of the classification process and various decisions within the process
have been subjective. The classification has as much as possible been based on theory
regarding landscape quality. If another researcher designed a landscape classification
process based on the same theory there may be some similarity in the resulting
classifications. However, there is a lack of such theory so this is unlikely to be the
case. The implementation of this classification is, however, totally objective. Automation
requires totally explicit instructions, and these have been described comprehensively in
this thesis. Does the classification produce a hierarchical
classification? Yes, and this has been demonstrated (refer
to Figure 6.3). Is the classification flexible so as to cope with new
interests and developments? The classification is flexible because it is modular. The
landscape attributes are assessed independently. If one attribute was discovered to be
unnecessary, it could be easily dropped from the classification. Similarly, if an
attribute was deficient or absent, it could be further developed or added. What also makes
this classification flexible is that it is totally explicit. People can see how it works
and can improve on it. In this classification, many critical parameter settings have been
used in functions, and definitions. These settings can be easily changed. For example, a
3000m search radius was often used. This could be easily changed if cognitive research
discovered this value to be deficient. A user friendly program could incorporate a menu
interface whereby the user selects appropriate settings, thus enabling the classification
to be very flexible. In section 5.4, the flexibility of the classification was
demonstrated by producing many different outcomes reflecting different conceptual models
of landforms. This can be done for the whole landscape classification process. The
classification can also combine different attributes of various generalisation as
demonstrated in section 0. Does the classification recognize seasonal or cyclical
change? Such change should be consistent within a class. If a
class changes in some areas with seasons, while in other areas it does not, then there are
perhaps deficiencies in the classification. The attributes most affected by seasonal
change are vegetation and water. Naturalness and landforms do not have cyclical changes.
Most of the vegetation and water classes within the classification generally change
consistently, however, there are some exceptions. Some exotic tree species are evergreen
while others are deciduous. The deciduous species can change quite dramatically with
seasons. The same can be said for some indigenous species. With the current vegetation
databases that are easily available, it is not possible to distinguish accurately between
evergreen and deciduous species. Perhaps individual forest companies could provide
coverages of their own forests containing this information. It can be generally said that
most forests in New Zealand are evergreen. The need for generalisation may make it
impractical to distinguish deciduous and evergreen, and also distinguish between
indigenous and exotic. Agricultural landscapes also change seasonally and this will be
inconsistent for some classes. The classification does not distinguish between crops and
pasture. A field that is being used for growing crops will change in appearance with the
seasons differently to a field in pasture. It is not possible to incorporate this
distinction with the currently available databases in New Zealand. If this information was
available, the generalisation issue may make it impractical to distinguish these classes. Water bodies also change seasonally - river flows change,
lake levels change, and coasts can be rougher at different times of the year. This again
can be inconsistent within a class. An obvious reason will be that these components are
affected by climate, which in turn is affected by topography. Some rivers are snow fed,
while others are not, and therefore flow differently during the spring thaw. The
classification, in its present state, does not consider these subtleties. It is probably
possible to incorporate them in an automated classification with present technology and
knowledge, and even more likely in the future when databases hopefully become more
sophisticated. However, one needs to question whether it is useful to include this
additional information when it is necessary to generalise. 6.4.2 Specific landscape classification criteria Does the classification incorporate landform,
vegetation, naturalness, and water? Yes, the classification was designed to do so. The more
important question is whether relevant components of these four main attributes have been
incorporated? This is difficult to say because landscape content category research has
tended to produce results only at a generalised level, and has not provided much insight
into how these main attributes should be further classified. Moreover, this research has
not been New Zealand based. The classification is limited more by our understanding of
landscapes then by GIS capabilities. This thesis has demonstrated the power and
flexibility of GIS for classifying attributes of landscape. When the important landscape
components have been decided upon and substantiated by content category research, then GIS
will probably be able to incorporate these in a landscape classification. This is the case
with landform components as there is now a body of research that has investigated
automated landform classification. Vegetation components are usually already identified
and are available digitally to a detailed level so can therefore be easily incorporated in
an automated classification. The same can be said for components of water. However, with
naturalness, automation appears limited by the complexity of the available databases. Some
classes of naturalness, such as tourism, mining, and heavy industry, cannot be identified
adequately with the current databases available in New Zealand. Is the classification based on the general public's
perception of landscape attributes? The classification attempts to classify from a general
public's perspective by using appropriate levels of generalisation. Since the general
public's perspective is not homogeneous, the classification is hierarchically structured
so that it can be used for a range of different perceptions. The classification therefore
addresses this criterion, even though it is not exactly known how the public perceives
landscapes. Is the classification based on an overall impression
of an area perceived from a distance, and does it involve generalisation and composition?
All the components included in the classification have
been of large enough size to be seen from a distance. This has been done by generalising
or grouping various subtleties so that together these groups are easily visible from a
distance. The classification not only incorporates compositions of the four main
attributes, but also incorporates compositions of individual components within these main
attributes. It also expresses the actual degree of composition (for example, 20% mountain,
50% hill, 20% scrub, 50% remote). The result is that the classification has the potential
to identify an astounding number of unique composites - 6.7 X 1017. Does the classification recognize landscape as an
experience from a multiple of perspectives obtained from movement and exploration? The classification uses focal mean functions
to incorporate movement and exploration. Manual classification techniques
have tended to confine analysis of areas to the visual extent of the neighbourhood,
while this automated approach has considered areas that are both visible
and not directly visible. A neighbourhood focal mean functions can consider
all the components in a set neighbourhood that are likely to be encountered
through movement and exploration and thereby contribute to the landscape
impression. The problem with considering movement and exploration is that
the extent and behaviour of these are not known. In this classification
a 3000m radius has been commonly used. Is 3000m appropriate? Theoretical
understandings provide no answers for this. It could also be argued that
exploration is not consistent over an area. People follow roads, and walk
on established paths (though their visibility is not confined to these).
Also, should areas far from the point of analysis be considered equally
as closer areas? To develop an automated classification that considers
these aspects, may be possible but would be complicated. Areas can be
classified in terms of accessibility to paths and roads. As discussed
in section 4.2.4., annuluses and kernels
can be used with focal functions, and these can be used to weight different
cells by the distance from the central cell. However, before such avenues
are researched, the usefulness of the normal focal function should be
first ascertained as this may be adequate. Only further research that
uses this classification process will answer this. 6.4.3 GIS errors As mentioned in section 3.5.2,
because GIS is particularly powerful with spatial information, errors
can be easily propagated. It is therefore necessary to assess these errors
and to ensure that the classification is not invalidated by them. These
errors can be grouped into three types - database errors, computational
errors, and logical errors. These have been called GIS errors but in fact
they can also be an issue to manual approaches. Since perceived landscape is a fuzzy entity
it does not permit precise measurement. If the boundary of the classes
was changed 200 metres, it would not make too much difference, as it is
not exactly known where the boundary should be anyway. This fuzziness
has been demonstrated in section 5.4.2.,
and the notion of entropy and agreement models have been used to address
this issue. When considering error, it is important to also consider the
error that is acceptable. With landscapes, there is quite a bit of leeway.
The amount of leeway appears to have never been discussed in the literature.
To determine this figure, requires research that compares results from
the use of several different classifications based on different spatial
extents. For the time being, this figure is very arbitrary. It is assumed
for this thesis that it is about 1000m. 6.4.3.1 Database errors Common sources of error are associated with data quality.
These have been classified as positional error, attribute accuracy, and spatial
resolution. They can result from data collection, data input, data storage, data
manipulation, and data output (Aronoff, 1991, and Bernhardsen, 1992). Data quality is a major issue within GIS mainly because
GIS uses many different databases (Chrisman, 1991). These databases can be easily shared
between different users of GIS and can be easily manipulated. Such is the ease of data
sharing and manipulation that it can be very difficult to determine the history of a
database and its level of accuracy. Positional accuracy With positional accuracy, the databases are adequate. The
accuracy of DOSLI's 1:250,000 topographical database is 150m for 90% of un-generalised
points (Newsome, 1995). For Landcare's vegetation database, the accuracy is 200m (Newsome,
1995). The accuracy of the Ministry of Forestry databases, which are digitised from a
1:250,000 base map, is unspecified. For the digital chart of the world, the accuracy is
specified, in the DCW metadata, as 7100m horizontally and 2000m vertically for the study
area. The first two databases were the more important databases for the classification,
and 200m is well within the 1000m assumed acceptable error limit. It should also be considered that mapping agencies
specify positional errors for a particular degree of certainty. For example, the DCW is
based on the ONC map series, which is used for airplane navigation. The mapping agents,
when specifying the positional error of these maps, had to consider the end use of these
maps and safeguarded themselves (against lawsuits) by specifying high error intervals.
There is not a vertical inaccuracy of 2000m in the New Zealand part of the DCW database,
however, this is what has been specified. With landscape classification, a high degree of
certainty is not required because human life is not at risk. It is therefore questionable
whether the positional accuracy of the databases specified by their publisher is relevant
for assessing the error of the resulting landscape classification, which needs
considerably less certainty. Spatial resolution The spatial resolution of the database, for the purposes
of this study, refers to the minimum size for polygons, or the minimum distance between
lines and points. Many databases have been obtained from hard copy maps, which can only
present a certain amount of information. If a map presents too many roads and polygons, it
soon becomes unreadable. If there are too many structures within a certain area, then
these are generalised to one structure. A minimum polygon size, and distance between lines
and points are used. For DOSLI's 1:250,000 topographic database, this is specified
comprehensively (DOSLI, 1984) and varies for different objects. The smallest polygon size
is 0.1 ha. All rural roads are recorded because they are not close together in reality.
Structures have been generalised if these were too close. Landcare's vegetation database
contains polygons 0.05 ha in size, however the spatial resolution is specified by Newsome
(1995) as 500 ha. The spatial resolution of these databases is adequate for the
classification process because the classification usually uses 20% as a minimum presence
within a 282 ha (3000m radius) neighbourhood. Unless accompanied by a sufficient number of
small polygons, small polygons will not make a significant difference. Also, detail in the
databases may be lost during vector to raster conversion since a 500m cell size was used.
This is a computation error and is discussed later. Therefore even if the databases had
higher spatial resolution, it is unlikely that this would make a difference to the
classification outcome. Attribute error Attribute error is concerned with whether a cartographic
entity (polygon, line, point, or cell) in a database is labelled correctly. These errors
may be present because of cartographic error, or because the database is out of date. All
useful databases will have attribute errors because of the need to generalise a complex
reality. If an area has mostly forest but also has some grassland, a useful representation
of this would be forest, which would not be entirely correct. Very general labels may be
used to reduce attribute error (eg. to call the above area vegetation), but might not be
useful. Attribute error is related to spatial resolution. For the purpose of landscape classification, the accuracy
of the 1:250,000 topographic database was sufficient for most entities. This database was
current in 1990. Some changes may have occurred since then but would not be significant.
If any major attribute errors existed, they would have been brought to DOSLI's attention.
The only exceptions to this are the tracks, mines and structures layers. Tracks were
missing in the more remote areas, and it would have been useful if more specific labels
had been used for the mines and structures layer (as discussed in section 4.2.2.). One may question the attribute accuracy
of Landcare's vegetation database. It was derived mainly from the Land
Resource Inventory and field work done before 1981. Also, considering
the nature of vegetation and recent agricultural and afforestation initiatives,
many labels will be incorrect. This was discussed in section 4.2.2
The Ministry of Forestry databases should be reasonably
free from attribute error as they were developed in recent years. Supermap2 and the DCW
were only used to identify towns and their populations. Supermap2 was derived from the
1991 census, and the DCW was revised in 1991 for the coverages that overlap the study
area. Considering their limited use in the classification process, their attribute error
would not affect the outcome significantly. 6.4.3.2 Computational errors Considering the number of computations that can be
implemented with a GIS and the degree of complexity of these, there is potential for error
to accumulate and become significant. Often with user friendly GIS interfaces, it is easy
to instigate a function but not know precisely how that function works and what
calculations are involved. With many GIS functions there is a considerable amount of
generalization and interpolation involved, and it is possible for the user to be oblivious
to this. Perhaps the most significant computational error in the
classification is associated with the conversion of vector data structures to raster data
structures. Vector coverages are usually a more precise way of representing landscape
components, however, they are more difficult to spatially analyse than raster coverages.
It would be very difficult to classify landscapes using only vector coverages. The effects
of vector to raster conversion have been mentioned with regard to the spatial resolution
of databases. The effect of this operation is dependent not only on the spatial resolution
of the databases, but also on the geometry of the polygons and lines, and on chance. With
the vector to raster conversion of polygons, the polygon class that contains the greatest
area within a cell will become the class assigned to that cell. A long narrow polygon,
which might occupy a large area, may be lost as neighbouring polygons might contain more
area for each cell. Whether this happens depends also on how the grid overlays the polygon
coverage, which is fairly random. It is possible for a polygon to be lost if it is just
less than four times the cell size, but the chance of a grid dividing a polygon into
exactly four equal size parts is small. Each part would be less than half the size of the
cell size, and could be lost if its allocated cell was shared with just one other polygon.
With vector to raster conversion of lines and points in
ARC/INFO, vectors are represented in a raster coverage by the overlapping cells. If there
is more than one line or point that overlaps a cell, then the majority class (based on
length of line or number of points) is allocated to that cell. In the classification
process, only single class vector coverages were converted to raster coverages, therefore
not too much attribute detail was lost. However, the vector to raster conversion only
recorded the presence or absence of a class, so if there were two lines or points of the
same class within a cell, the result was the same as if there was only one of these.
Whether vector information is lost depends on how the grid overlays which is usually
fairly random. The vector to raster conversion also spatially generalises vectors. For
example, a twenty metre wide road in a vector coverage can become a 500m road in a raster
coverage. However, this was not a problem as all components exerted a spatial influence of
at least 3000m. The other main source of computational
error that exists in the classification is terrain interpolation. The
representation of a terrain surface using TIN created some obvious errors
with the landform classification process developed by Dikau et al. (1991),
as discussed in section 5.3.2. There is always
error associated with TINs because they interpolate and this can be significant
in flat areas that have neighbouring relief. It was necessary to ensure
that the new landform classification process was sensitive to this error.
This was done by not having too many class that were dependent on subtle
changes in slope. It is difficult to ascertain the error
associated with the slope measurements because it is not valid to compare
the results with manually calculated results. As discussed in section
5.3.2, the method used for calculating slope
affects the resulting slope calculations, and GIS slope functions use
a different method to manual slope measurements. 6.4.3.3 Logical errors A type of error that has been associated with landscape
assessment is, for the purposes of this study, called logical error. Hamill (1989)
provides an alarming account of these errors that have persisted for a long time within
landscape research and have been largely uncontested. He used Leopold's method as an
example (Leopold, 1969). Here, numbers were used incorrectly (spurious numbers) as they
were assigned arbitrarily to denote different classes. These numbers are therefore nominal
numbers, but they were used in mathematical operations as if they were cardinal numbers.
The results of these operations were not only meaningless but varied depending on what
number was assigned to which class. Dearden (1980, p.52) also comments on this persistence of
error. He states that "these measurement techniques contravene the theories of levels
of measurement by using nominal or ordinal scales of measurement and then employing
standard arithmetic procedures, such as multiplication and addition. In these
circumstances, the methods become invalid." Lowenthal (1978, p390) sums it up nicely;
" adding together landform and landuse, panoramic
and historical features is like summing apples, oranges, bacon and peppercorns." Within a GIS environment, it is often necessary to
represent words by numbers, especially within raster coverages. The mathematical
manipulation of these numbers is very easy within a GIS and care is required to ensure
that this is done appropriately. Logical errors do not exist within the classification,
although it may appear so. Many operations in the classification are not arithmetical.
They combine coverages rather than add coverages. An "and" operation was used
instead of a "+" operation. With "apples" and "oranges", the
effect of a combine operation is to get a new coverage with a class of "apples and
oranges", not some spurious attribute value. When arithmetical operations were used
they were done within a class rather than between classes. For instance, the focal mean
function was applied to single theme coverages. There is nothing wrong with saying, "
an apple plus an apple equals two apples". Manual and automated approaches can be compared by
discussing them in relation to the general and specific criteria. The automated approach
has been subjected to this and now it is appropriate to do the same with the manual
approaches. Concerning general classification criteria, manual
classifications may be mutually exclusive, exhaustive, hierarchical, and able to
incorporate seasonal or cyclical changes. However, they are generally not explicit and
cannot produce repeatable results that are independent of the observer. This is because
they tend to be intuitively based. Some degree of replication may be possible if people
have had similar training and some objective criteria are used. However, in relation to an
automated method, they are no match. At a national or international scale, there is
unlikely to be repeatability with manual methods. It is arguable whether manual methods
are easily understood and applied. How can a method be understood and applied if it is
intuitive? This must lead to confusion as practitioners seek confirmation on the exact
nature of landscape components and composition. Also, manual methods must be very time and
resource consuming as many landscape need to be directly observed. Manual methods may be
flexible because they are vague, but how can new understandings be gained when it
uncertain how the method was implemented. Regarding specific classification criteria, the manual
approach does have some credibility, but it is debateable whether this is more than the
automated approach. There is no doubt that the direct perception of landscapes in the
field will give a better indication of the nature of a landscape than a computer. However,
a landscape in one area has to be classified in relation to the landscapes in other areas.
The practitioner therefore has a massive amount of information that needs to be
considered. It is questionable whether this can be done manually over a large area, such
as the size of New Zealand, or the world. Manual classification of large areas may not involve
field visits to view the entire study area. Representations of reality, such as maps, are
often used instead, but the information required is unlikely to be all on one map. Can
humans analyse effectively several maps at a time to get an overall impression of
landform, vegetation, naturalness, and water? This must be a tedious, and challenging
task. It is likely that the manual approaches that use maps have to separate landscape
into main attributes, like the automated approach, to make the classification manageable. One task that manual methods do well is
the recognition of pattern. For example, the recognition of a valley floor
(relief-flat-relief) can be done easily manually. However, to do this
automatically, involves considerable processing. The same can be said
for the identification of compositions. As commented in section 5.3.2,
some other patterns are probably more effectively recognised using manual
methods, for instance the topographic patterns associated with conical
volcanoes. Concerning the notion that landscapes are experienced
from movement and exploration, the manual method has been deficient in the past. Most
manual methods appear to use direct visibility, and according to the criteria this is
inappropriate. Even if manual methods did incorporate exploration, there would not be the
resources available to fully explore the whole study area. It would be necessary to rely
on maps produced by surveyors that have already done the exploration. The question then
returns to whether humans or computers are more effective at analysing maps? Manual methods require that the practitioner divide the
study area into analysis units (usually areas of visual enclosure) before analysis of
landscape character begins. This is necessary to make the task manageable. With
automation, the GIS divides the study area and the analysis unit is kept very small in
comparison. The dramatic subdivision of the study area (into many cells) effectively
enables the analysis of landscapes from point perspectives. This is more appropriate and
would replicate landscape perception, which is also done from many different points. In conclusion, it can be said that the automated approach
is superior in terms of general classification criteria because it is explicit and
repeatable. Regarding the specific criteria, both manual and automated approaches have
their pros and cons. Overall, because explicitness and objective repeatability are
essential ingredients for a classification, then automation is a significant improvement
for landscape classification. 6.5 Applications 6.5.1 Frame of reference The most significant application for a
landscape classification is as a frame of reference for communication
within landscape research. This has been discussed in section 2.6.
From the results, such as presented in 2, it can be seen that if someone
was researching landscapes within the study area, then they could use
the classification in a variety of ways: for description, mapping, and
inventory purposes. Firstly, a particular location can be described by
the class within which it is located, or a region can be described by
the predominant landscape classes that exist within it. Secondly, all
localities with certain landscape characteristics can be located and mapped
using the classification. Thirdly, inventories can be created showing
areas and numbers of occurrences of landscapes satisfying certain conditions.
With GIS, it very easy to generate information on the total area of different
classes within a specified region. Within the GRID module of ARC/INFO,
a value attribute table is generated that counts the total number of cells
that exists for each class. The area of each class can then be calculated
by multiplying this number by the area of each cell (25 ha in this study).
This has been done in 2 for the study area. It can be said, for example,
that at generalisation level L3-V3-N3-W3, there is 54,850 hectares of
mountain / indigenous scrub / remote landscape within the study area.
When the classification is in vector format, the number of occurrences
of different classes can be calculated in ARCPLOT simple by selecting
them. The area of single polygons can also be easily ascertained. For this type of communication to be effective on a
national or international scale, standardisation is required for the different levels of
generalisation, as well as the labels that describe each class. The labels used in 2 have
been chosen because they describe the actual class. This is a useful coding system but can
distort people's interpretation of the class. Each class should be interpreted with the
underlying explicit definitions of these classes. By using a descriptive label, people may
be inclined to use their own conceptual definitions of these labels to interpret these
classes. Non descriptive codes could be used, for example A5R6, but these would make it
difficult for people to become familiar with the classification. 6.5.2 Determining uniqueness As discussed in section 2.6.,
uniqueness has been used for assessing the value of different landscapes.
However, there is not a clear relationship between landscape value and
uniqueness. If a landscape is unique and considered ugly, then it is of
little value. However, uniqueness can make an average landscape important,
or a beautiful landscape extremely important. Information on uniqueness
is therefore sought after by landscape practitioners. Strictly speaking, if something is unique then it is the
only one of its kind, therefore, something is either unique or it is not. However, whether
something is unique is often expressed on a relative scale, as implied in the phase,
"quite unique". The concept of uniqueness has evolved although the term,
"rarity" might be more grammatically correct. Uniqueness appears to be on a
scale from absolutely unique to very common. Uniqueness of a class can be expressed using the
percentage of the total area a class occupies. This would depend on size of the analysis
area (scale), and also on the level of generalisation in which the landscape is perceived.
With the classifications that have been generated, it is now possible for people to be
explicit about these two considerations, which in turn, could lead to more constructive
debates within planning courts. Inventory statistics can be divided by the total area of
analysis to give an impression of the uniqueness of that class within the study area.
People can now question whether that level of generalisation is important, and whether the
extent of the study area is relevant. As discussed previously, the level of generalisation
can be easily changed. The same can be said for the extent of the analysis area. In the above example, the study area was the extent of
analysis. With GIS, it is easy to change this extent of analysis by setting an analysis
mask. This has the effect of "cutting" the coverage to the required extent. For
example, if areal statistics were required just for the Banks Peninsula region, then a
coverage that just shows the extent of this region could be used to set the extent of the
analysis. A new classification coverage of Banks Peninsula can then be generated by simply
entering the command, "coverage (Banks Peninsula) = coverage (study area)". This
new classification coverage of Banks Peninsula will automatically have a value attribute
table with areal information for each class. Uniqueness information can therefore be
generated for all levels of scale - local, regional, national, etc. However, the analysis
has to be confined to the extent of the available classification, which at the moment is
only for the study area. With GIS, it is relatively easy to change the extent of
the analysis in incremental steps, and for each incremental step generate uniqueness
information. To start with, an analysis window of one cell can be used, which can be
located over the point of interest. Areal information for a particular class would be
either zero or 25 ha (for 500m cell size). The coverage that defines the extent of
analysis can then be expanded by one cell in all directions, thereby, generating an
analysis window of nine cells. Areal information can then be generated again. This
procedure can be repeated hundreds or thousands of times automatically until the size of
the analysis window is more than required. Such analysis would be reasonably quick for an
area that was the same size as the study area. The resulting information would enable, for
a particular class, the uniqueness to be plotted in relation to scale. Instead of asking
the question, "is this class at this location unique at this scale?", it is now
practically feasible to ask the question, "at what scale is this class located here
unique?". However, before these questions can be answered, the question, "what
is unique?" would first have to be answered. The same incremental uniqueness analysis can be
calculated for different levels of generalisation, for a given extent of analysis.
Uniqueness can be calculated for generalisation level one, and then for level two, and so
on, until all generalisation levels have been considered. The uniqueness of a class, for a
particular location, can then be plotted against generalisation, and the question that can
be asked is: "At what level of generalisation is this class, at this location, for
this extent of analysis, unique?". By combining this information with the analysis of
scale, as described above, a very interesting model of what is unique would develop.
However, before it is worthwhile to develop these models, which are not pushing GIS
technology to its limit, it is first necessary to agree on a landscape classification,
with its different levels of generalisation. Landscape variety has been used for assessing
landscape quality by the Ministry of Works and Development (1983 and 1987),
however, the validity of this has not been proven (Arthur et al., 1977).
Whatever the case, there is a demand for information on landscape variety.
Variety can be defined as the number of unique classes within a given
area. In the past, landscape practitioners have not used quantitative
techniques for assessing variety but have used a more intuitive approach.
With GIS, variety can be calculated by using a focal variety neighbourhood
function (Berry, 1993). This is similar to other focal functions used
in this study but assigns the number of unique classes that exist within
the analysis window to the central cell that is being processed. The analysis
window can be of any size or shape. 4 shows the effect of such a function
on the landscape classification developed in this study. The analysis
window was a 5000m radius circle. If the size of the NAW changes, then
so will the variety. Intuitive means for assessing variety are implemented
very subjectively, while with GIS, once variety has been defined and a
landscape classification agreed upon, then it can be implemented objectively.
Since GIS uses an explicit definition of variety, then this definition
can be questioned and developed, as our understanding of the nature of
landscapes improves. Figure 6.4 also
shows the effect of generalisation on variety. As expected, variety is
very dependent on this. Obviously, the more detailed a classification
and the greater the search radius, then the greater the number of classes
that are likely to exist within a given area and therefore the more variety.
This figure demonstrates that when practitioners are considering variety,
they also need to consider the level of generalisation. 6.5.4 A basis for further manual classification If automatic landscape classification is considered
inadequate, a hybrid of automatic and manual classification could be considered. As
discussed in section 0, both the manual and automatic approaches have their advantages and
disadvantages. Perhaps if manual and automatic methods were considered together, then
these disadvantages may disappear. As discussed in section 6.4.4, manual approaches appear
more appropriate for identifying particular patterns or shapes, such as, conical volcanos.
It could be considered that some individual landscape components can be identified better
manually, but the spatial extent and composition of these components can be calculated
better using GIS. Such a hybrid approach is feasible, however, the landscape components
would need to be made available in digital format. 6.5.5 A means for understanding landscapes An interesting spinoff from trying to classify landscapes
automatically is the increased understanding that is obtained about the nature of
landscapes. To develop the classification presented in this study, required a considerable
amount of "simulation" that considered the effects of using different
components, spatial extents, and other parameters. By assessing the effects of these, the
importance of different components and parameter settings became apparent. The ability to
perform hundreds of classifications is a major advantage of automation and is useful way
of exploring the nature of landscapes. Automation also requires detail about landscapes to be
explicitly addressed. With manual approaches, based on intuition, many details have not
been considered explicitly. As a result these details have not come out into the
intellectual arena and been openly discussed. It has been argued previously that the best method for
validating a classification is to use it. If an unacceptable discrepancy becomes apparent
within a class, then this may demonstrate that the nature of landscapes is more complex
than the classification portrays. In this way, a classification evolves and an increased
understanding of the nature of landscapes is obtained. Automation facilitates this
evolution because a classification can be easily redesigned and reapplied.
6.4.4 Manual versus the automated approach
6.5.3 Assessing landscape variety
Automation has the advantage that a considerable amount of information on
various attributes and over a large area can be treated consistently. With manual
classification, it is doubtful whether a practitioner, or a team of practitioners, can
match this consistency. However, the exact boundaries of classes may be determined better
using manual approaches, especially if areas of visual enclosure are considered a better
basis for analysis than focal means. With visual enclosures, the boundaries of classes
often correspond with the crest of ridges. While with focal means, this is not so. What is
best depends on how one thinks landscapes are perceived. It has been argued in this thesis
that landscape perception is derived from movement and exploration, which is affected but
not completely restricted by ridges. If visual enclosures are considered more appropriate
for landscape classification, an automatically generated classification could form the
basis for a classification, which could then be altered manually. The boundaries could be
manually edited to ensure they match catchment boundaries. This could be done in digital
format using GIS editing capabilities.
