This paper was presented by Stan as a key note speech at the International Conference on Modelling Geographical and Environmental Systems with Geographical Information Systems in Hong Kong, 1998.
1. Background and Context
2. Design of a Synoptic Prediction System (SPS)
2.1 Basic design objectives
2.2 Basic System
2.4 Stages in Operationalising a SPS
3. Assembling, aggregating, and estimating the data
3.1 The spatial interpolation problem
3.2 Creating a common spatial database
3.3 Estimating the Population distribution
3.4 Climatic Data
3.5 Other Environmental Data Sets
4 Results of Modelling Land-use change.
4.1 Now Land-use Modelling
4.2 Future Land-use Modelling
4.3 Assessing the impact of change
There is a growing likelihood that global climatic change will soon start to have a visible and increasingly fundamental environmental and socio-economic impact within the next 25 to 50 years across many parts of the world. In some regions the effects may well be unnoticed or are irrelevant or are well within the capacity of existing ecosystems to cope; however, in other regions the environment is far more fragile and susceptible to serious problems. In these regions it is possible that even small changes in the climate may be sufficient to cause a major impact on the environment and socio-economic systems that relate to it. The research challenge is to identify a plausible way of making reasonable predictions of climatic change impacts on land-use for 25 to 50 years hence as a possible basis for raising awareness and creating a framework for action. The hardness of this challenge should not be underestimated, but equally there is an increasing urgency to know something of what may be happening to our world in the medium term future so that thinking and strategic planning may be proactive rather than purely reactive.
From a methodological perspective there has been very little research performed on predicting future land use patterns on a local let alone national or European scale. Most computer models that exist are generally concerned with only very limited aspects of the problem; for instance, regional employment change or population dynamics. An exception is the work by the CLUE Group (see de Koning et al (1997), Veldkamp and Fresco (1996, 1997), Verburg et al (1997)). The CLUE modelling framework (The Conversion of Land Use and its Effects) is based on a multi-scale stepwise regression approach that attempts to analyse and model land-use and land-use change as a function of socio-economic and biophysical factors at an aggregate spatial scale for China, Ecuador, and Costa Rica. The model is a linear continuous time simulation at a fairly coarse spatial scale; ranging from 7.5 km2 for Coasta Rica to 32 km2 for China. For the European Union (EU) we feel that something far more sophisticated is needed. The basic requirements are that; the level of spatial resolution is sufficiently detailed to be useful, the key driving factors and process mechanisms that link human and physical environments are explicitly incorporated, and that the model is driven by climatic and environmental change. Additionally, the model should seek to make good use of the research and model results produced by other teams involved in the MEDALUS III.
The principal justification for a modelling approach is that a more traditional mapping of indicators will probably fail to incorporate many of the most important process variables, such maps are static, they can only realistically be produced for small case study regions (inhibiting the formation of an objective global overview), and the data and methodology would be hard to standardise so that like may not be compared with like. This is not a criticism that mapping land degradation and of environmentally sensitive areas for this or that part of the EU is not useful, only that this may be the wrong approach if the objective is to provide a synoptic overview of global climate change impacts on land degradation across the whole Mediterranean climate region of the EU.
Additionally, if scientific research on land degradation processes is to have a major political and public impact commensurate with the academic and theoretical importance of the subject then a means needs to be found, however imperfect, to translate the scientific research into a form that non-technically sophisticated decision-makers and a non-technically sophisticated public can understand and appreciate. GIS provides a good map based communications medium but also what are needed are consistent results that are mapable, based on the best science at the present moment in time, and which show what global climate change impacts are on agricultural land-use.
Section 2 outlines the structure of a Synoptic Prediction System (SPS) which is designed to meet this challenge. Section 3 describes how the various data components were assembled. Section 4 discusses the operation of the system to make both nowcasts and forecasts of future land-use patterns. Section 5 presents some more general conclusions.
In operationalising Figure 1 the choice of input variables is restricted to those available for MEDALUS III research which reflect "obvious" processes. Figure 2 outlines the SPS in greater detail. The available variables are not ideal, but then probably no one knows what would be ideal in this context. Nevertheless Figure 2 shows that we are using the most obvious variables. Table 2 gives the full list of 18 predictors used in the modelling which is described in Section 3.
A further word of caution. The results will critically depend on the quality of the inputs, and although this has not yet been quantified, we believe the current SPS uses the best available data even though it is deficient. The aim is to provide broad brush forecasts which are not necessarily accurate but which offer a synoptic view of the likely impacts. The SPS simultaneously demonstrates what is needed to model the process of land degradation as well as indicating the likely effects of global climate change on land use. If the results are what is required then doubtlessly their accuracy can be improved. If critics do not like this approach, then let them demonstrate that they can do better given the same objectives, the same data restrictions and research resource constraints that apply here. What is proposed here captures the very essence of a GIS based approach to modelling environmental systems from a geographical perspective.
Section 3 describes Steps 1 and 2. Section 4 covers 3, 4 and 5 whilst step 6 is yet to be completed.Step 1: Assemble the data for a common EU wide geography for the present day (circa 1991).
Step 2: Obtain forecasts for these variables for 25 years hence (notionally 2023) and 50 years hence (designated 2048) for the same geography.
Step 3: Construct neural nets to model the relationship between climate (temperature and rainfall), soil characteristics (permeability, texture, fertility, parent material), biomass, elevation, population densities, and other socio-economic variables to predict contemporary land-use patterns.
Step 4: Compute estimates of land-use for 2023 and 2048 by using forecast values of the inputs to the neural nets and also investigate different climatic change scenarios.
Step 5: Create maps of the impact by comparing these forecasts with predictions for the present day.
Step 6: Consider modifying the forecasts and the predictions to reflect knowledge expressed as fuzzy rules and repeat Step 5.
Using GIS most available environmental data for the EU can probably be manipulated into a regular grid orientated at a spatial resolution of approximately 1 km2. A grid was selected as the framework in which to store, manipulate, link and map the data since it offers the greatest flexibility in aggregating upwards and can yet still provide a realistic representation of regional or local variation provided the grid cells are sufficiently small. A geographical latitude-longitude projection was chosen as a compromise given the traditional problems in map projections regarding the representation of distance, direction and area distortion of the data caused by the curvature of the earth. A 1-decimal-minute (1 DM) resolution (which is roughly equivalent to a 1 km2 for most of the EU) was selected as providing the most appropriate and probably the best possible level of spatial resolution that was practicable for this research.
A common spatial framework for all the available climatic, environmental, and socio-economic data is an essential pre-requisite before any modelling can be attempted. This is not a trivial task as typically the climatic data are produced at a coarser scale than the socio-economic which is itself far coarser than that at which many physical models have been applied. Aggregation upwards is fairly trivial but interpolation from coarse to finer levels of spatial resolution is far more problematic and error prone yet this is an unavoidably essential activity that needs to be mastered before much progress can be made. Virtually every data source involved in the SPS and described in Figure 2 (see also Tables 2 and 3) had a unique set of problems associated with it and necessitated various GIS operations and sometimes modelling applications to create a common scale data base.
The first data set to be processed was the Global Land 1 km Base Elevation source data (GLOBE). This provides a grid of 0.5 DM resolution cells in a geographical projection based on the World Geodetic System 84 datum. It was imported into ArcInfo. The grid was aggregated to a resolution of 1 DM where each 1 DM cell was assigned the mean value of the four 0.5 DM cells from which it was composed. Since the average height above sea level values of this height variable are normalised by the neural net program the sum or some other composite combination of these values could have been used. There are several reason for selecting to use the mean; compared with the mode and median, it is easier to compute since there are four values being summarised, the loss of data precision is less than if just a random selection of one of the four values was assigned, and finally the resulting 1 DM values are still standard distance units of height above sea level whereas they would not be if the sum had been used. The grid was clipped to a size of 2205 rows and 2568 columns which covered the whole of the EU and most of the rest of Europe. The GLOBE data itself did not need to be projected, but all the other source data based on different projections had to be transformed. There are various ways of doing this, they all suffer from problems and there is a trade off depending on the nature of both the non-nesting aggregations between the projections and the nature of the spatial variable (whether it is density, distance or direction related).
Gridded night-time lights source data was imported into ArcInfo, converted into a polygon coverage and projected from its original Goode Homosline projection into the required geographical projection using the projection capabilities of the software. The projected polygons were intersected with a polygon coverage which coincides with the grid in the chosen 1 DM spatial framework. A value was calculated and attached to each small intersected polygon by dividing the night-time lights intensity value by the area of this small polygon. The intersected polygon values within each 1 DM grid polygon were then summed and the resulting coverage was converted into a grid.
Similarly, the Surpop 0.2 km source data was imported into ArcInfo, converted into a polygon coverage, projection, then intersected with a polygon coverage which coincides with the 1 DM chosen spatial framework, and again the intersected polygons were assigned proportions of the population depending on the area of the intersections. As before the intersected polygon values within each 1 DM grid polygon were summed and the resulting coverage was converted into a grid. The total population in the source data was compared with the total population in the transformed data to ensure that they were not significantly different. The values in the transformed data were generally not integer values, although this was not a problem for the neural net, and an integerised version was created for mapping purposes.
The Bartholomew digital map data was manipulated into various grids in the 1 DM spatial framework reflecting either; the location of, distance from, or density of geographical features. To begin all the various map layers were imported into ArcInfo and mapped using ArcView. Geographical features which appeared to be consistently defined across the EU and whose location, proximity or density were believed to be correlated with population density were manipulated into location, cost-distance and density layers respectively. Cell values in location layers were either 0 or 1 depending on whether the cell lay mainly inside or outside the location of a selected geographical feature. For the distance layers the spatial analyst module of ArcView was used to assign the proximity of each cell to selected geographical features. Creating the density layers employed a point-in-polygon routine and the grid module in ArcInfo. The density of a selected spatial variable or geographical feature was calculate at various spatial scales and some results from certain resolutions were aggregated to coarser spatial scales where their values appeared to spatially correlate with population density. The coarser resolution grids were then disaggregated in a simple fashion where each 1 DM cell was assigned the value of the larger cell in which it was contained. Using a weighted linear function all the density grids relating to a particular theme were combined at the 1DM resolution, these combined grids were again mapped to examine the correlation with population density.
Finally population related variables were included even though they occur at coarser levels of geography; for instance, population densities at NUTS 3 scale and RIVM's estimates at 10 km2 scale data. The values of the 1DM cells were simply assigned the closest value of these coarser data units. The purpose was to provide a multi-scale contextual element.
The aim, therefore, is to use widely available digital map derived summary variables that are probably related in some non-linear surrogate fashion with population density; for instance, road network density, distance to the nearest train station, the location and size of settlements, height above sea level, etc. The full list is given in Table 2. There is also some external knowledge that can be imposed on the results; in particular, areas known to be uninhabited (e.g. sea or lakes) can be set to zero whilst known population counts in NUTS 3 regions can be used to constrain the predicted values.
The basic idea, therefore, was to use the 1991 Surpop census population surface to build a neural net based spatial interpolator to relate population density to a selected set of predictor variables. For each 1 DM cell the values of the variable chosen to model the population densities were concatenated into a large file of vectors from which a randomly selected small training data set of 10,000 1 DM cells was created, together with the associated Surpop counts. The small number of training cases reflected a desire not to produce nets that only worked well in the UK. Ideally, the training data should have been based on data for multiple countries in the southern EU but no small area census data were available to us from which Surpop-like estimates might be provided. As a result there is a risk that population digital map surrogate relationships are different in the southern EU (i.e. different lifestyle) but there was little that could be done to reduce this source of uncertainty due to an absence of data. The EU really does need to organise its basic data resources to a far better degree than is current. It is really most unsatisfactory that even NUTS 5 (equivalent to UK ward level) resolution data are not available throughout the EU and that data copyright and ownership prevents access to high resolution data even for those applications where the results are of potential public benefit.
A variety of simple feed forward perceptron networks were applied. Tests indicated that those nets with a single hidden layer of 25, 50, 75, or 100 neurons were out performed by a net with two hidden layers each with 20 neurons in them. The neural net training used a hybrid approach: first an evolutionary optimiser was used to find a good solution and then this was fine tuned using a conjugate non-linear optimisation method. The trained network weights were then applied for the rest of the data across the EU. The NUTS 3 population totals from Eurostat were used to constrain the predictions of the 1 DM cells in each area. Errors were analysed using the Surpop data in the UK at the 1 DM scale but also at the NUTS 5 scale in Britain and Italy (the only two countries for which we had these data). Tests were also made of the likely improvement if NUTS 5 data had been available across the EU.
The results appear to be remarkably good; see Figure 3. The predicted surfaces correctly pick up the main features of the population distribution of the EU even if there is a slight loss of peakiness. It was surprising how well these surfaces matched reality given the nature of the input data and with further post-processing to add lumpiness could further improve the estimates. Use of accounting constraints based on NUTS 5 data appeared to make little difference to the results. Finally, forecasts for 2023 and 2048 were produced by using available EU forecasts for 2023 in NUTS 3 areas, and our own for 2048 (as no official ones existed) to constrain the estimates.
Estimates of potential biomass were provided by MEDALUS III colleagues researching at the University of Leeds. This is the output of a model which translates temperature and rainfall data into measurements of expected or potential Biomass. The present potential biomass model is fairly primitive as it does not take into account factors like the height above sea level or soil type and the output used thus far is at a relatively coarse level of resolution as it has been built from 30 DM resolution monthly temperature and rainfall data. Nevertheless, the 30 DM resolution potential biomass data was imported into ArcInfo and interpolated into the desired 1 DM resolution using the interpolation capability provided in the spatial analyst extension of ArcView. It is used here mainly to provide a contextual variable.
The other inputs concern a set of broad land-use categories. There were two land use source variables attached with the soils source data which relate to dominant and secondary land use classes. These classes are derived from a satellite imagery, there is considerable uncertainty regarding the class of secondary land use but the dominant land use classification is believed to be relatively accurate. The dominant land use variable was thus selected as the target land use for which the neural network trained to classify. Prior to this dominant land-use was aggregated into four broad categories; arable, trees and orchards, wasteland and others and each in turn was used as a dependent target variable to train the contemporary land use classifier. Each cell was assigned a value 0 or 1 depending on whether it belonged or not to the dominant land use class which was being modelled. The classification could be greatly improved by assigning values for each cell, based on the original satellite data, which give the probability of each cell belonging to a particular land use class; see Moody et al (1996) and Carpenter et al (1997).
However, it is important not to overlook the deficiencies. To be frank the results are broad brush and can be criticised on the following grounds:
On the other hand, the SPS does have some good points, in particular:
the market mechanisms and agricultural subsidy levels which may well link (somehow) environmental change to the socio-economics to produce a land-use response are implicit rather than explicit and assume a continuation of the present; the neural net model results could be improved if better quality climatic and environmental data were available; the uncertainty in the outputs has not yet been made explicit; there is a mixture of inputs with very different levels of data uncertainty and forecast reliability; global climatic effects are equivalent to a shift in the boundaries of agricultural capability; it assumes technology remains more or less the same; and, the land-use categorisation is very crude.
Of course the forecast predictions will be wrong! The hope is that when aggregated to an appropriate level of geography they will not be so wrong as to be useless. The aim is to raise awareness and to communicate the possible impacts on land-use 50 years into the future. It is a straw man! Let those who dislike these results demonstrate how with existing science they can do better. It is also a challenge for those who like the results, the onus being to improve them by reducing the uncertainties in the inputs and enhancing the modelling that was used. There is nothing in this paper that could not be improved either by the availability of better data, improved forecasts, and more key variables; or by the input of more effort to enhance the modelling. However, given current data, current knowledge and current science it is difficult to see how we could do much better. We wanted to create a need for land-use forecasting models that incorporate climatic, environmental, and socio-economic variables. We have chosen to meet this goal by outlining a practical system, however imperfect. If the results outlined here are at all useful, then maybe the resources needed to improve them will be forthcoming. Meanwhile we would argue that our results are unique in that they are all that exists right now so the principle of caveat emptor should be applied. The results are the first of their kind and really only serve as a benchmark and a preliminary test of methodology. All in all, the SPS appears to provide a useful framework for assessing the possible impacts of climatic change on land-use by linking all the various components in a novel and interesting way.
it is the first attempt to predict 2023 and 2048 land-use changes linked to forecast climatic change; there is a linkage of physical - environmental and socio-economic aspects; the same methodology has been consistently applied across the southern EU; it is a brave attempt to make broad brush land-use impact predictions for 50 years ahead; it offers a different but useful approach to assessing the possible impacts of climatic change on land-use by linking all the various components in a novel and interesting way; it has produced broad brush results relatively quickly which can be updated as new and improved outputs from socio-economic and other environmental models become available; it is difficult to see how the challenge could be done in any better way at present; and, the results are understandable and should help focus the political debate about how to handle these desertification problems by putting them into a pan-EU context.
Further details of the research can be found on the www at http://medalus.leeds.ac.uk/SEM/home.htm
Table 1. Advantages and disadvantages of neural networks
|universal approximators||computationally intensive|
|equation free||may require long training times|
|highly non-linear||choice of architecture is subjective|
|promise of good performance||depends on training data|
|handle hard to model problems||black box technology|
|automated||conveys little knowledge|
2. Variables used to create European population surfaces
|Digital Elevation Model 2|
|Night time lights intensity at 1 km scale 3|
|Distance from nearest built up areas 1|
|Distance from nearest canal 1|
|Distance from nearest international airport 1|
|Distance from nearest national park 1|
|Distance from nearest river 1|
|Communications network density 1|
|Motorway and dual carriageway road network density 1|
|Main and minor road network density 1|
|Railway network density 1|
|Distance from extra large towns 1|
|Distance from large towns 1|
|Distance from medium sized towns 1|
|Distance from small towns 1|
|Location of built-up areas containing extra large town centres 1|
|Location of built-up areas containing large town centres 1|
|Location of built-up areas containing medium sized town centres 1|
|Location of built-up areas containing small town centres 1|
|Location of named settlements and built-up areas 1|
|Regiomap population density at NUTS-3 4|
|Tobler's pycnophylactic population density based on NUTS-3 5|
|RIVM's population density at 10 km scale 6|
|Surpop Great Britain Census target population density 7|
|Variable Label and Data Source||Description|
|Location of soil type 1 11||This includes the following soil classes; cambisol, chernozem, luvisol, vertisol, plaggensols.|
|Location of soil type 2 11||This includes the following soil classes; rendzina, gleysol, phaeozem, fluvisol, kastanozem, histozol, andosol.|
|Location of soil type 3 11||This includes the following soil classes; arensol, ferralsol, ranker, planosol.|
|Location of soil type 4 11||This includes the following soil classes; acrisol, podzoluvisol, greyzem, podzol, solonchak.|
|Location of soil type 5 11||This includes the following soil classes; solonetz, xerosol.|
|Location of soil type 6 11||This includes the following soil classes; lithosol, regosol, rock outcrops|
|Soil quality 11||Physical properties of the soil were indexed in terms of their limitations or restrictions for agricultural capability and combined to produce a crude measure of soil quality.|
|Potential biomass 10||Estimated potential biomass model output at 30DM resolution.|
|Average temperature in Spring 9||Average monthly air temperature in March, April and May.|
|Average temperature in Summer 9||Average monthly air temperature in June, July and August.|
|Average temperature in Autumn 9||Average monthly air temperature in September, October and November.|
|Average temperature in Winter 9||Average monthly air temperature in December, January and February.|
|Average monthly precipitation in Spring 9||Average monthly precipitation in March, April and May.|
|Average monthly precipitation in Summer9||Average monthly precipitation in June, July and August.|
|Average monthly precipitation in Autumn9||Average monthly precipitation in September, October and November.|
|Average monthly precipitation in Winter 9||Average monthly precipitation in December, January and February.|
|Digital Elevation Model 12||Height above sea level.|
|Population||1x10x10x1 Neural Network output.|
|Dominant agricultural land-use 11||The dominant agricultural land use categorised into the following groups; arable, olive groves and orchards, wasteland, and others.|
Burrough P, MacMillan R, van Deurson W (1992) Fuzzy Classification Methods for Determining Land Suitability from Soil Profile Observation and Topography. Journal of Soil Science, 43, 193-210.
Carpenter G, Gopal S, Martens S, Woodcock C (1997) Evaluation of Mixture Estimation Methods for Vegetation Mapping. Boston University Center for Adaptive Systems and Department of Cognitive and Neural Systems technical report CAS/CNS-97-014.
Clarke J, Rhind D (1991) Population Data and Global Environmental Change. International Social Science Council Programme on Human Dimensions of Global Environmental Change, Working Group on Demographic Data.
Deichmann U, Eklundh L (1991) Global digital datasets for land degradation studies: A GIS approach. United Nations Environment Programme, Global Resource Information Database.
Deichmann U (1996) A Review of Spatial Population Database Design and Modelling. United Nations Environment Programme, Global Resource Information Database, and Consultative Group for International Agricultural Research initiative on the Use of Geographic Information Systems in Agricultural Research.
de Koning G, Veldkamp A, Verburg P, Kok K, Bergsma A (1997) CLUE: A tool for spatially explicit and scale sensitive exploration of land use changes. Working Paper, Wageningen Agricultural University, The Netherlands.
Elvidge C, Baugh K, Kihn E, Kroehl H, Davis E (1997) Mapping City Lights With Nighttime Data From the DMSP Operational Linescan System. Photogrammetric Engineering and Remote Sensing, 63, 727-734.
Elvidge C, Baugh K, Kihn E, Kroehl H, Davis E, Davis C (1997) Relation between satelite observed visible-near infrared emissions, population, economic activity and electric power consumption. International Journal of Remote Sensing, 18 (6), 1373-1379.
FAO (Food and Agriclture Organisation), 1980, Natural resources and the human environment for food and agriculture. Environment paper No 1, Rome, Italy.
Goodchild M, Anselin L, Deichmann U (1993) A framework for the areal interpolation of socio-economic data. Environment and Planning A, 25, 383-397.
Kosko B (1994) Fuzzy Thinking: The New Science of Fuzzy Logic. Harper Collins.
McMahon M (1998) The home page of MEDALUS III: http://medalus.leeds.ac.uk/
McNeill D, Freiburger P (1994) Fuzzy Logic: The Discovery of a Revolutionary Computer Technology. Simon and Schuster.
Moody A, Gopal S, Strahler A (1996) Artificial Neural Network Response to Mixed Pixels in Coarse-Resolution Satellite Data. Remote Sensing Environment 58, 329-343.
Openshaw S, Turner A (1998) The home page of MEDALUS III project 9.1: GIS based socio-economic modelling: http://medalus.leeds.ac.uk/SEM/home.htm
Openshaw S, Perree T (1997) Methods for an intelligent continuous surface transformation of socio-economic data for Europe. Extended Abstract for 1st International Conference on Geocomputation, September 1997, Leeds, England.
Openshaw S, Openshaw C (1997) Artificial Intelligence in Geography. Wiley & son.
Palutikof J, Agnew M (1997) Statistical Downscaling of Model Data. MEDALUS III report from the Climatic Research Unit at the University of East Anglia.
Ross T (1995) Fuzzy Logic with Engineering Applications. McGraw-Hill.
Sweitzer J, Langaas S (1994) Modelling population density in the Baltic States using the Digital Chart of the World and other small scale data sets. Technical Paper, Beiher International Institute of Ecological Economics, Stockholm.
Tobler W (1979) Smooth Pycnophylactic Interpolation for Geographical Regions. Journal of the American Statistical Association, 74, 519-529
Tobler W, Deichmann U, Gottsegen J, Maloy K (1995) The Global Demography Project. NCGIA Technical Report TR-95-6.
Turner B, Moss R, Skole D (1993) Relating land use and global land-cover change: A proposal for an IGBP-HDP core project. Report from the IGBP-HDP Working Group on Land-Use/Land-Cover Change. Joint publication of the International Geosphere-Biosphere Programme (Report No. 24) and the Human Dimensions of Global Environmental Change Programme (Report No. 5). Stockholm: Royal Academy of Sciences.
van Veidhuizen J, van de Veide R, van Woerden J (1995) Population mapping to Support Environmental Monitoring: Some Experiences at European Scale. Proceedings of the Joint European Conference and Exhibition on Geographic Information, March 1995, The Hague, Netherlands.
Veldkamp A, Fresco L (1996) CLUE: a conceptual model to study the Conversion of Land Use and its Effects. Ecological Modelling, 85, 253-270.
Veldkamp A, Fresco L (1997) Exploring Land Use Scenarios, An Alternative Approach Based on Actual Land Use. Agricultural systems, 55(1), 1-17.
Verburg P, de Koning G, Kok K, Veldkamp A, Fresco L, Bouma J (1997) Quantifying the spatial structure of landuse change: an integrated approach. ITC Journal Special Issue: Geo-Information for Sustainable Land Management.
Warren A, Agnew C (1988) An assessment of desertification and land in arid and semi-arid areas. International Institute for Environment and Development (Paper No. 2), London: Ecology and Conservation Unit, University College.