During the last 10 years the University of Utrecht in the Netherlands
has pursued a continuing strategy concerning the development of methods
and software tools for spatial and temporal analysis. These tools
have been directly linked to the needs of environmental scientists working
in fields as diverse as radioecology, epidemiology, hydrology, soil science,
sedimentology and physical geography. The driving forces behind the
conceptual and software developments have often come from specific user
needs; persons responsible for the quality of the environment, the protection
of resources, or the avoidance of disasters have been increasingly turning
to information technology to supply them with the data, the tools and the
models to help them predict how landscapes may respond to natural or anthropogenic
changes.
Standard GIS tools (by which I mean those for dealing with digital versions of paper maps and remotely sensed data) have provided useful means of data storage, data retrieval and data visualisation, together with a specific, but limited set of tools for data analysis. Today, the standard spatial entities in many environmental databases are still the supposedly the simple points, lines or homogeneous map polygon whose attributes can be analysed with many logical and mathematical tools. Many current users of digital environmental data still have few ideas of spatial interactions and spatial-temporal change, or of variation and data quality, so the methods provided by standard commercial software packages suit their needs. This situation is reinforced because computer scientists responsible for software development often view environmental data as being similar if not identical to other kinds of spatial data, such as is encountered in utility applications or land ownership systems. So long as this is the situation, there will be little reason for vendors to provide methods of analysis for which there is little demand.
At Utrecht, we realised that in order to deal with many kinds of spatial-temporal problems there had to be new developments to supplement the spatial and temporal analysis methods provided by standard GIS packages. We chose to deal specifically with the following:
1. The modelling of dynamic processes in space and time
2. Geostatistical interpolation and simulation
3. Exploratory data analysis
4. Error propagation in spatial modelling
5. Multivariate indices of spatial patterns and spatial change, in
particular using methods of fuzzy logic.
6. Visualisation of spatio-temporal processes
7. Educational aspects of spatio-temporal modelling
1. The modelling of dynamic processes in space and time
Many environmental scientists working with dynamic processes (e.g. groundwater flow, erosion, runoff, etc.) have used GIS as a source of data which are downloaded to a model (e.g. MODFLOW). The model is run and results are returned to the GIS for display. This treats the model as a black box – it is fine if you can accept the way the model works and can supply the data it needs. On the other hand, if you want to change the model you need the source code and skills in computer programming.
Of course you can always write your own model in a language of your
choice, but not everyone is a skilled programmer or has time to put together
large amounts of code. So we realised that there would be many advantages
to creating a generic tool for spatial-temporal modelling. Such a
tool would make use of the command line interface common in raster GIS,
but would provide a higher level programming language in terms most scientists
could understand. A generic tool (a spatial-temporal version of MATHCAD)
would:
a) Make writing and modifying models easier
b) Standardise the model interface
c) Optimize links between commands, models and the database
d) Provide a sound basis for teaching and research.
The dynamic modelling tool is called PCRASTER (http//www.geog.uu.nl/pcraster).
It operates in raster mode and contains more than 150 spatial operations
drawn from the rich resources of map algebra, cellular automata, hydrological
routing, image filtering and so on. Additional routines can be supplied
by the user via a plug and play interface. The main developments
were done by Willem van Deursen (1994) and Cees Wesseling and have been
continually added to by Cees and his colleagues since. PCRaster is
now used by many government institutes and universities (from the European
Union to individual researchers) to supplement standard GIS. It has been
used in applications as diverse as the reactions of large river catchments
(the Rhine, Bramaputra) to possible climatic change, nutrient flows in
large catchments, soil erosion at scales from the Mediterranean to metre-square
plots, the modelling of deltas and river meandering, landslides, the dispersion
of plants and animals, predator-prey interactions and many more.
PCRaster also enables the user to view the results of spatio-temporal models
as 3-4D movie-like ‘draped’ displays so that the ways a landscape reacts
to the various processes can easily be seen. PCRaster grids
can be very easily exported to ARC VIEW. Information about the theory
and applications of PCRaster can be found in Wesseling et al 1996, Burrough
and McDonnell 1998, and Burrough 1998.
2 Geostatistical interpolation and simulation
Most spatial models that are run either in PCRaster or other formats require space to be discretised, either with regular cells (finite difference modelling) or defined entities or polygons (finite elements). In most cases data need to be collected from sparsely located points and then interpolated to fill the gaps. Conventional methods of interpolation are quick and dirty, and the methods of geostatistics provide a rich suite of tools for optimal interpolation of static spatial patterns. One main advantage is that geostatistical methods also give an indication of the quality of the interpolation and the errors associated with it.
Gstat, written by Edzer Pebesma, is a very comprehensive geostatistical toolkit that provides means for variogram estimation and fitting, and most commonly encountered forms of kriging interpolation including point and block estimation, simple, ordinary and universal kriging, indicator functions and stratification according to external criteria. Gstat uses the same spatial data format as PCRaster so that both the interpolated surfaces and the information about strata can be easily exchanged. Data input is via the well-known Geo-Eas format.
Gstat also includes methods for conditional simulation of spatial surfaces, which provides means for studying the role of either random or spatially coordinated errors in modelling. This provides Monte Carlo methods for following the propagation of errors in the dynamic PCRaster models. Some dynamic models with local interactions (e.g. river meandering or the modelling of alluvial fans and deltas) need a stochastic seed to get them started and Gstat provides ways of creating these randomised inputs.
More information about Gstat and how to get it can be found at gstat-info@geog.uu.nl
3. Exploratory data analysis
The provision of hyperlinked windows in statistical packages and in programs like ARC VIEW have greatly simplified the detection of errors in data and the amount of insights a user can get before carrying out complex analyses or modelling. At Utrecht we did a certain amount of work with John Haslett’s group in Trinity College Dublin on the addition of geostatistical analyses to his REGARD programme (Gunnink and Burrough 1996). Unfortunately the Macintosh software could not easily be transferred to Windows so developments in REGARD ceased. Today, programmes like Yves Pannatier’s VARIOWIN (not a Utrecht product) and S-plus, SPSS, etc. provide much useful exploratory data analysis tools that are easy to use in conjunction with PCRaster and Gstat.
4. Error propagation in spatial modelling
Gerard Heuvelink’s work on error propagation in spatial modelling (now published by Taylor and Francis – Heuvelink 1998) was a pioneering attempt to link Geostatistics and GIS in such a way that one could identify the different sources of uncertainty in the results of GIS models, and the magnitude of the contributions from each source. By linking this work to previous work on the optimisation of sampling networks by McBratney and Webster (1981) it is possible to carry out a cost-benefit analysis of different combinations of interpolation methods and data configurations (See Burrough and McDonnell 1998, Chapter 10).
5. Multivariate indices of spatial patterns and spatial change, in particular using methods of fuzzy logic.
Another major line of research at Utrecht has concerned the applications of multivariate methods for classifying spatial patterns. While most statistical packages provide factor analyses and numerical clustering tools, few provide methods for fuzzy classification. Imposing the rules of an existing fuzzy classification on mapped data is little more than a standard computational operation in GIS, but deriving an optimal, overlapping fuzzy classification from point data requires other means. In common with many other researchers in this field, we have used the methods of fuzzy k-means (see Burrough and McDonnell 1998, Chapter 11). We have demonstrated that in order to achieve coherent patterns of multivariate groups it is not only essential to have a good clustering in data space, but also a strong spatial correlation structure (as expressed by the variogram).
Fuzzy membership values computed from point data can of course be interpolated
by geostatistics to space filling grids, and the interpolation imposes
a certain degree of spatial continuity. We have used this method
for multivariate classification of soil, geochemical data and crop yield
variations (multiyear results of different crops with harvesting aided
by GPS). When data are taken from continuous surfaces such as DEMs,
the inputs are already spatially correlated and variogram analysis is not
necessary. Recently, together with Bob MacMillan and John Wilson,
Pauline van Gaans and myself have used PCRaster (for the derivation of
other DEM attributes like slope, plan and profile convexity, ridge proximity,
together with simulation modelling of derived drainage networks and fuzzy
k-means to create stable classifications of landforms. Applications
range from a 150 ha site in Alberta, Canada to a 8000km2 plus area of the
Yellowstone National Park. Independent tests of the classification
demonstrate stability with extension to neighbouring areas, and also in
terms of ecological properties of the derived units.
6. Visualisation of spatio-temporal processes and patterns
Having clever methods of modelling and analysis is no good unless you
have the appropriate tools to present the results to the user. Cees
Wesseling and Victor Jetten have developed a range of display tools, based
on games routines, for the dynamic display of spatial-temporal models.
7. Educational aspects of spatio-temporal modelling
We have been teaching GIS and Geostatistics to both undergraduates and graduates for many years. Since 1996 we have also been teaching the methods of dynamic modelling to external and internal researchers and PhD candidates, and since 1997 to final year students so that they can use these methods in their major fieldwork studies. The results have been very encouraging indeed. An ongoing project is to create a series of virtual landscape tools in which students can explore how landscapes may react to short term and long term changes in control parameters. This enables them to follow processes that may take place very quickly (e.g. raindrop splash) or processes that take millions of years (e.g. tectonic uplift). By combining PCRaster models with Gstat functionality in a user-friendly shell (currently Powerpoint) it is easy and effective to link models to explanatory text, figures, photos and video. Another ongoing project being led by Derk Jan Karssenberg and Cees Wesseling is our contribution to an EU-sponsored Distance Learning project, which is being coordinated by Joao Ribeiro da Costa at the New Technical University in Lisbon, Portugal.
8. New developments.
New developments include the provision of a complete Windows interface,
a 3D-4D database structure to deal with issues arising in sedimentology
and erosion, improved multivariate methods, better visualisation tools,
etc.
References.
Burrough, P.A. and A. Frank (Eds.) Geographic Objects with Indeterminate
Boundaries Taylor and Francis – 1996.
Gunnink J. & Burrough, P.A.. 1997. Interactive spatial analysis
of soil attribute patterns using Exploratory Data Analysis (EDA) and GIS.
In: M. Fischer, H.J. Scholten and D. Unwin. (eds). Spatial Analysis
in GIS.. Taylor and Francis (1996). 87-100.
Burrough, P.A.. Environmental Modelling with Geographic Information
Systems. In: Innovations in GIS 4. Editor Zarine Kemp, Taylor
and Francis 1997 , pp 143-153.
Burrough, P.A. and J. Swindell. 1997. Optimal mapping of site-specific
multivariate soil properties. In: J.Lake, G. Bock and J. Goode (Eds) Precision
Agriculture: spatial and temporal variability of Environmental Quality,
Proc: CIBA Foundation Symposium 210, John Wiley and Sons, Chichester, 1997.
208-220.
Burrough, P.A. & R.A. McDonnell. Principles of Geographical
Information Systems (2nd edition) Oxford (February 1998), 333p., reprinted
August 1998.
Burrough, P.A.. Dynamic Modelling and GIS, Chapter 9, In: P.Longley
et al.(Eds) Geocomputation: a Primier. Wiley 1998, pp165-192
Burrough, P.A., P.van Gaans & R. Hootsmans. Continuous Classification
in Soil Survey: Spatial Correla-tion, Confusi-on and Boundaries Geoderma
77: 115-35. 1997.
Burrough, P.A., P.F.M. van Gaans and R.A. MacMillan. High-resolution
landform classification using fuzzy k-means (Journal of Fuzzy Sets
- in press 1998/9:
Burrough, P.A., P.F.M. van Gaans, J. Wilson and A.J. Hansen. Fuzzy
k-means classification of digital elevation models as an aid to forest
mapping in the Greater Yellowstone Area, USA (in preparation)
Pebesma, E. and Wesseling, C.G. 1998. Gstat: a program
for geostatistical modelling, prediction and simulation. Computers
and Geosciences 24: 17-31.
Heuvelink, G.B.M. 1998. Error Propagation in Environmental Modelling
with GIS. Taylor and Francis
Van Deursen, W.P.A. 1994. Geographical Information Systems
and Dynamic Models. Netherlands Geographical Studies No. 190.
Formerly Senior Lecturer in Spatial Analysis and Soil Science, Wageningen Agricultural University, NL (1980-1984), Senior Research Scientist, Dutch Soil Survey (1976-1980), Lecturer in Soil Science, University of New South Wales (1973-1976), Research Scientist, Land Resources Division, Tolworth, London (1970-1973).
Research includes:
- applications of geographical information systems for all kinds
of environmental resource survey, both within western countries and in
developing countries
- International GI issues, mainly in Europe, in connection with
EUROGI, and more recently AGILE. (Association of Geographical Information
Laboratories in Europe)
tel: 030 253 2766
fax: 030 254 0604
email: P.Burrough@frw.ruu.nl