7.1 Can GIS classify landscapes?
This thesis demonstrates that GIS can classify landscapes using characteristics that the theoretical literature appears to consider important.
The classification process can be summarized as follows:
1. The selection of the four landscape attributes.
2. For each landscape attribute:
a) The identification of important components from existing NDDBs using attribute and spatial generalisation. Indented coast and landforms had to be conceptualised using complex routines.
b) The determination of the spatial influence of each component using a focal neighbourhood mean function.
c) The identification of classes using complex conditional queries on the spatial influence information.
d) Generalising the classes to six levels by grouping the
classes using simple attribute generalisations.
3. Overlaying the landscape attribute classes for each level of generalisation and using the unique combinations as landscape classes.
The classification of landscapes using GIS has considerable advantages over traditional manual methods. The most important advantage is the ability to be explicit and repeatable with complex definitions. Automated classification is also flexible because operational definitions can be easily changed and regions can then be easily reclassified. Once an automated process is developed, it is also considerably quicker to apply than manual methods.
A major issue with developing any classification is validation. A set of criteria using general classification criteria, and specific landscape criteria appear useful for this. Operational errors, and a comparison with previous methods have also been considered. Specific landscape criteria consider whether the classification incorporates the important characteristics of landscapes that affect quality. The classification has generally met these criteria. The classification can now be challenged in terms of whether the criteria are adequate, and/or whether these criteria were adequately applied. Because landscape theory is a "theoretical vacuum" (Appleton, 1975b, p.2), it was only possible to include generalised characteristics in the landscape criteria. These criteria need to be met in order for the classification to be valid, however, they are not sufficient to judge whether the classification is actually valid. Validation needs to be completed through independent research.
This research highlighted many deficiencies in the current theoretical understanding of the nature of landscapes, particularly in relation to their composition. These deficiencies have been stated previously in a more general manner by Steinitz (1993), and Mitchell (1993), however, GIS forces one to be very specific about them. It can be considered an advantage of automation that, because GIS needs everything to be explicit, all details need to be addressed.
"[A] computer-based information system serves to highlight deficiencies in theory, not to hide them, and its use must be seen as an aid to better understanding" (Duffield and Coppock, 1975, p.141).
Specific deficiencies identified in this research relating to the nature of landscapes were:
. What are the exact landscape components that are important to landscape perception and the nature of their contribution?
. What are the appropriate distant decay functions from a given point for each of the landscape components?
. What are the important component compositions?
The lack of landscape theory makes it difficult to substantiate many decisions made in the classification process and therefore the classification is subjective. This does not mean that this classification is inappropriate. For theory to develop, a classification is required to act as a frame of reference. A classification needs to be developed as best as possible with existing theory. Where this is deficient, assumptions should be made. As theory develops, the classification then needs to be revised if assumptions are proven to be incorrect, and it is in this way that a classification evolves.
It appears that GIS can provide an effective research tool for developing landscape theory. A range of possible landscape classifications can be developed and these can then be tested through psychophysical and cognitive research. In this way, an increased understanding of the nature of landscapes can be obtained and the above questions might be answered.
Entropy and agreement models can be used to cope with the gaps in theoretical understanding and the fuzziness of landscapes. This was demonstrated and discussed in section 5.4.2. regarding landforms. It was concluded that the use of an agreement model is more appropriate than the use of entropy for this purpose. This model provides information on the certainty of a classification enabling a researcher to select areas for future investigation. If a researcher wants to research a landscape that has a high consensus (or perhaps a low consensus) over its identity then information on this can be obtained. For example, an evaluation study using psychophysical methods will probably use high consensus areas. However, a researcher may be interested in low consensus areas to develop understanding of landscape character.
Focal neighbourhood functions, in particular the focal mean function, are valuable tools for landscape classification and offer an alternate approach to conventional methods. They can effectively be used to calculate the spatial influence of components from thousands of different points, since relatively small cells (500m) act like points. These spatial influence measurements can then be grouped into classes. Following Zube, Sell, and Taylor's (1982) emphasis that movement and exploration are important parts of landscape perception, it has been argued in this thesis that the focal functions can incorporate these. Focal functions are therefore theoretically and practically preferable to a visibility function.
A significant part of this study has been developing a method for identifying landforms as this attribute had not been conceptualised in existing databases for the study area. This is a reasonably complex operation but, once developed, it is an effective means for classifying macro landforms. A significant problem with automating landform classification is the measurement of slope. It appears that when GIS is used for measuring slope, the results will be different to manual slope interpretations. This is because methods for manual slope measurements have not been explicitly stated, in terms of scale and location of measurements. An explicit method can be applied with GIS, whereby the scale and location of measurements are stated. Since GIS and manual slope measurements give different results, then slope thresholds used in manual classification are likely to be inappropriate for automated classification. When seeking a GIS measured slope threshold to distinguish flat and non-flat areas, a slope threshold of four percent, rather than Hammond's eight percent, was the most appropriate. A four percent slope threshold measured with GIS had 91% agreement with a seven degree threshold based on the LRI's manually measured slope information. Other problems identified with existing automated landform classification methods were the mixing of macro and meso objects, and the way the spatial influence of relief is determined. These problems can be resolved, and this was demonstrated in a new landform classification developed in this study. It appears that automating a manual classification process, such as Hammond's, may be a good place to start when developing an automated classification but modification is often needed, and improvements can be made to existing manual processes.
This thesis has also revealed options for future research
that could improve the existing classification. As previously stated, more cognitive
research on the nature of landscapes is needed, but there are also GIS options that could
be investigated. These being the use of annuluses and wedges for specifying the extent of
the NAW, the use of kernels for incorporating an appropriate distance decay function for
landscape components, the identification of particular landforms, such as conical
volcanoes, and the use of more complex databases as they come available, eg. the 1:50,000
topographic database. The use of visibility functions could also be considered for
providing extra information, but it is doubtful whether they will be more appropriate than
focal neighbourhood functions. Perhaps they could be used in combination.
Duffield and Coppock (1975, p.146) made the following comment concerning their computerised landscape assessment package:
"Perhaps the primary deficiency for landscape assessment lies in the system's inability to cope with the spatial composition of landscape, as opposed to its resource content. It is not unique in this failure; indeed, nearly all existing procedures of assessment of landscape have proved incapable of dealing with this vital aspect of the appeal of landscapes. Clearly the appreciation of landscape is primarily aesthetic and derives as much, if not more, from the spatial relationship of visible resources as from their mere presence in the scene."
GIS can express complex spatial relationships of
landscape components, especially if annuluses and wedges are used for defining the NAW.
However, the problem is that there is no agreement on what spatial relationships are
important. Before more complex spatial relationships are expressed with GIS, it is
necessary that our understanding of the nature of landscapes is improved, and this needs
to be based on an existing classification.
7.2 Implications for databases
This research used national digital databases, in particular, DOSLI's 1:250,000 topographic database and Landcare's vegetation database. Since NDDBs are relatively new and are still in the process of being developed, it is worthwhile to comment on their worth and possible improvements.
This study has demonstrated that these databases are particularly useful for complex spatial analysis of large areas. These databases in themselves contain a lot of conceptual information, for example towns, roads, etc. However, with spatial analysis, additional concepts can be identified that are useful for resource planning. This has been demonstrated by identifying not only variety and uniqueness, but also certain components of naturalness and landform. In the context of landscape description, it is important that these databases are seen as more than raw material for cartography, and that their true worth is realised.
Despite the fact that digital databases are of such worth, this research has also demonstrated that at the moment there are lost opportunities with digital databases that need to be realised. Perhaps the greatest opportunity that is being lost is that they are not being fully utilized. This is related to access. Based on the experience encountered in this study, access to databases is currently inhibiting their use, and the greatest barrier to access is their cost.
Further opportunities can be gained from digital databases by making them larger and more complex. If this was the case, then they could still be easily analysed. The databases used in this study were fairly large (some were about 20 megabytes). However, this did not pose a major problem with the hardware and software. These databases were significantly reduced in size when converted to raster coverages with a 500m cell size. They could then be easily manipulated and duplicated without lengthy processing times or significant hard disk storage problems. The hard disk space used in this study was not more than 600 megabytes, and this included all the postscript and graphic files generated for presenting maps. The implications of this are that even larger and more complex databases can be used.
Currently, the information in topographical databases in New Zealand is generally limited to the amount of information on hard copy topographical maps. GIS can cope with far more information than this, even over a large area. If there is too much information, this can be easily generalised, but if there is not enough then this can severely restrict its application. The building layer is an example where more information would be useful. Currently, there is very little attribute information associated with this layer. If there had been more, then additional subtleties relating to naturalness could have been included. This study would have benefited if this layer had identified different types of structures rather than just having a single general category called "structures". Attribute information, such as the size, age, and even the number and type of occupants could have enabled a tourism class to be identified. Enhancements to the urban layer, and the mines layer would have also been beneficial. In this study, several different databases had to be used to get information on urban areas. The actual populations were obtained from Supermap2, medium size towns from the DCW, and the large urban areas and very small towns from the topographical database. It would have made analysis easier if all this information had been available within one database. As discussed in section 4.3.6, the mines layer was deficient because it contained only one class of mines. From this it is not possible to distinguish whether the mine is underground, open cast, in use, or abandoned.
General purpose databases (secondary) should contain as little generalisation as possible, and instead leave generalisation in the hands of the users of the databases. GIS is very capable of generalisation. The databases, however, have often been already significantly generalised by cartographers. It is recognised that generalisation is necessary for developing these databases, however, where practically feasible this should be kept to a minimum for users who can use powerful GIS. Inconsistency in the databases is the GIS user's nightmare. Generalisation when applied unevenly will result in inconsistency in the database. Automated generalisation is generally consistent in its application. If a user can apply their own generalisations then they can be sure that this is done consistently, and to the required level. A range of databases with different degrees of generalisation may, however, be appropriate for the benefit of others.
Standardisation of spatial and attribute data within NDDB is absolutely critical and has been highlighted by this study. Rule based automation needs to use consistent databases. Otherwise, it can produce spurious output. If inconsistent databases are used, then automatic processes need to be considerably complex to cope with the diversity of possible data. The simple solution is to ensure that the databases are standardized. It appears that this concern is already being addressed. Land Information New Zealand (LINZ) has released a series of publications specifying the standards that they will use (LINZ, 1985, 1987a, and 1987b). This covers standards for labels, geographical referencing, and measurement and inclusion of area size. Standardisation is also being attempted at an international scale (Murcott, 1995). These initiatives are important for the development of automated geographical abstraction. They, however, need to be applied by all agents that are providing digital databases, and not just the main mapping agents, such as DOSLI. For example, if place names had been standardised between different databases then this would have saved considerable inconvenience. Preferably DOSLI and Supermap2 should be using the same place names.
Databases should also be available in raster format. This
study has demonstrated the power of spatial analysis using a raster format. It is doubtful
whether such analysis could be done by using only vector coverages. Mapping agents are
supplying mainly vector coverages but it is raster coverages that are the most useful for
spatial analysis. This did not pose too many problems for this study since a vector to
raster conversion function was available. However, this used a powerful GIS that is not
available to many GIS users. Many cheap GISs are raster based, therefore, databases should
be made available for these systems. A DEM is difficult to obtain accurately from a vector
contour coverage as this study demonstrated. It would be preferable if mapping agents
supplied a range of DEMs with different cell sizes then the task of creating an accurate
DEM from contours would not have to be repeated by different users.
7.3 Implications for GIS
This study has demonstrated that complex geographical abstractions can be implemented within GIS to produce coherent, meaningful results. GIS can use a range of representation techniques that express classification, association, generalisation, and aggregation. This has enabled GIS to express complex structural geographical meaning. The challenge to do this for landscapes was mostly with regard to incorporating generalisation and association. Landscape classification is very much an exercise in generalisation as there is a considerable amount of digital data on different landscape components. Spatial and attribute generalisations using relatively complex conditional queries appear appropriate to bring the complexities of reality down to a manageable level. Simple attribute generalisation based on the grouping of classes is particularly useful for developing a workable hierarchy of levels whereby information relating to classes at specialised levels can be easily linked and applied to classes at a general level. The use of focal neighbourhood functions has been particularly useful for expressing association between components, which in turn has enabled compositions to be identified.
An important method for experimenting with different operational definitions is simulation. GIS with its associated macro languages can simulate complex processes thousands of times within a short period. This has been valuable for accurately "tuning" operational definitions of complex geographical concepts.
With hardware and software, there is always room for improvements. As hardware becomes faster, there is more processing demanded because new computer intensive applications become apparent. Since the process developed in this study was relatively quick, it became feasible to apply the classification several times using different definitions. This consumed significant amounts of CPU time, and a faster hardware platform would have made this task easier. The software could be improved by the removal of limitations on the number of coverages that can be used within a function, for example the majority function in GRID. Also, the removal of limitations on coverage sizes, such as, the number of arcs and size of attribute tables, would also be advantageous. Such limitations might be appropriate when people are experimenting with GIS and are not too sure of outcomes, but can be unwanted when large processing tasks are required. With the development of large national and global databases, extremely large tasks will be expected from GIS.
This study has demonstrated the power and ease with which GIS can manipulate and analyse information traditionally obtainable from hard copy maps. Most measurements that can be obtained manually from a map can now be done automatically. The automation of cartographic analysis significantly enhances the geographer's analytical opportunities. Extensive regions can be analysed, and this analysis can be quite complex. This has been demonstrated in this study with automated landscape classification. The ability of GIS to analyse large areas means that Geographers can realistically, quantitatively examine issues at a national scale, rather than be confined to regional or local scales because of practical constraints. For many environmental issues the national scale is important, and the landscape issue is an example of this. It can be argued that global analysis is also important, and it is only a matter of time before this type of analysis will be available.
This study has revealed how GIS can provide a platform from which a comprehensive landscape classification can evolve. Such a classification can be used for effective communication between landscape researchers, and could contribute to the development of consensus among researchers on landscape issues.