Forecasting global climatic change impacts on Mediterranean agricultural land use in the 21st Century

Abstract

This paper describes the development of a prototype Synoptic Prediction System (SPS) for the European Union (EU) funded Mediterranean desertification and land use project (MEDALUS III). The prototype SPS was designed to forecast the possible impacts of global climatic change on agricultural land use patterns across the Mediterranean region of the EU. Designing such a system was a challenging task because many of the theoretically desirable data sets were either unavailable or did not exist, whilst significant uncertainties were apparent in the available data. Additionally, process knowledge was woefully deficient as virtually all the principal mechanisms for linking the dynamics of the climate and physical environment with the associated socio-economic systems were poorly understood. In order to make the best predictions of the geographical impacts of climatic change for around 75 years hence we developed an integrated, synoptic, GIS, scenario based modelling approach as we believed it was the best available option. So, the paper describes the development of a prototype Synoptic Prediction System (SPS) which employs a mix of GIS, neurocomputing, and fuzzy logic technologies to attempt the almost impossible yet important task of forecasting agricultural land degradation risk under various climate change scenarios. The paper focuses on methods developed to enrich the available data, the quantitative approach to modelling and forecasting land use using neural networks, and the fuzzy logic based translation of the modelling results into land degradation terms.

Key words

Acknowledgements

1. Background and Context

This paper is based on research undertaken by the authors on a major European Union (EU) funded project concerned with Mediterranean desertification and land use. This project, MEDALUS III consisted of a wide range of topics aiming to describe, analyse and model various aspects of climatic and environmental change at different spatial scales in the Mediterranean region, see McMahon (1998), MEDALUS III final report (1999). The overall objective of our topic in this project was to devise some means by which to integrate socio-economic dimensions to the MEDALUS III physical models of agricultural land degradation. This task could have been tackled in various ways, our strategy was to develop a Synoptic Prediction System (SPS) that would model relationships between climatic, physical and socio-economic variables in order to translate potential climate change scenarios into likely agricultural land use impacts. One of the tasks that still remains is to package up the various modelling components into an integrated spatial decision support system that could be used by EU politicians, land use planners and the public to raise awareness of the problems, and help encourage proactive mitigation to alleviate them. For scientific research on land degradation processes (like MEDALUS III) to have political and public impacts commensurable with the academic and theoretical importance of the subject, then the scientific research needed translating into a decision-making form that politicians and a non-technically sophisticated public can understand and appreciate. GIS provided a good map based communications medium to visualise the results across the EU, but it arguably the most important thing is to base results on consistent definitions and the best available science.

There is a growing likelihood that global climatic change will soon start to have a visible and increasingly fundamental environmental and socio-economic impact in the 21^st century across many parts of the world. In some regions the effects may well be un-noticed or are irrelevant as they are well within the capacity of existing ecosystems to cope; however, in other regions such as Mediterranean climate zones the environment could be far more fragile. In the Mediterranean it is thought likely that seemingly small changes in climate maybe sufficient to cause a major impact on the environmental and socio-economic systems that relate to it. The ongoing research challenge is to identify and test ways of predicting climatic change impacts on agricultural land use 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 was little published research focussing on predicting future agricultural land use patterns on a local let alone national or EU scale in 1998. Most computer models that did exist seemed only to be concerned with one or two socio-economic aspects of the problem; for instance, regional employment change, or population dynamics. An exception was 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) was based on a multi-scale step-wise regression approach that attempted to analyse and model land use and land use change as a function of various socio-economic and biophysical factors. The model was a simple linear continuous time simulation operating at a fairly coarse spatial scale; ranging from 7.5 km² for Costa Rica to 32 km² for China. For the EU something far more sophisticated was needed. The basic requirements of the research in 1996 were that:

1. the level of spatial resolution was sufficiently detailed to be useful,
2. the key driving factors and process mechanisms that link human and physical environments were explicitly incorporated, and
3. that the entire system was driven by climatic change scenarios.

The principal justification for a modelling approach was that a traditional map based analysis of indicators had a high risk of failing to recognise the importance of many of the process variables. Additionally, such maps are static and realistically could only be produced and subjectively analysed in detail for small case study regions. Further still, the data and methodologies employed in such map based analyses were not standardised so invariably like was not being compared with like. Although the mapping of land degradation and environmentally sensitive areas for this or that part of the EU state was thought to be useful, it was believed to be insufficient for our objective of providing a synoptic overview across the entire EU Mediterranean climate region.

Section 2 outlines the structure of a Synoptic Prediction System (SPS) designed to forecast agricultural land degradation. Section 3 describes how the various data components were assembled. Section 4 describes the operation and results of predicting contemporary and forecasting future agricultural land use. Section 5 describes the fuzzy logic based translation of the forecast land use changes into land degradation risk indicators and their subsequent combination with other risk indicators to form scenario based consensus. Finally, section 6 presents some general conclusions and suggestions for further research.

2. Building a Synoptic Prediction System

The initial objective of integrating socio-economics into MEDALUS III physical models evolved into the task of developing a prototype Synoptic Prediction System (SPS) designed to estimate plausible impacts of climatic change on agricultural land use across the entire Mediterranean climate region of the EU. Designing and developing the system was challenging because many of the theoretically desirable data sets were either not available or did not exist whilst there were significant uncertainties in all the available data. Additionally, system process knowledge was woefully deficient as virtually all the principal mechanisms for linking the complex dynamics of the physical environment and climate with the associated socio-economic systems were poorly understood. In essence the underlying and most basic problem was that physical and climatic process models were far better developed than any of the models of socio-economic systems related to agricultural land use. If you asked the question: 'What is the likely effect of climatic change on the hydrology of a specific river catchment or on the speed of erosion of a particular hill slope?' - then methods existed that could indicate broad answers. If you asked the question: 'What is the likely impact of EU agricultural subsidies on the crops growing in the fields on this hill slope or on the agricultural land use of this catchment?' - then there were no existing models for the current situation let alone models which could forecast the effects in the medium to long term! Some kind of novel computer modelling methodology had to be devised to satisfy the forecasting objective, cope with the deficiencies in the available data, and manage without good process knowledge.

2.1 A prototype SPS

The desire was to:

1. Devise a GIS based model to predict contemporary agricultural land use using; climate variables (temperature and rainfall), environmental variables (relating to soil, biomass, elevation etc.), and socio-economic variables (including population density).
2. Forecast future agricultural land use under various climatic change scenarios.
3. Translate the predicted agricultural land use changes into surfaces of land degradation risk in order to generate synoptic land degradation forecasts.

2.2 Problems

The principal problem with the prototype SPS outlined here is that the relationship between the climate, environment and land use is not conditioned by the laws of physics or chemistry or biology but involves many ill defined human factors. In particular, the relationship is (at least) mediated and affected by available technology, market mechanisms, historical tradition, inertia, culture, demography and various economic factors such as land price and subsidy rates. Most of these aspects were effectively invisible to the reported application of the prototype SPS as they were not explicitly present in any of the available data that were used. It would of course have been better run a process model faster than real time into which crop prices, EU agricultural subsidies, irrigation practices, and each farmer's micro-behaviour were input. However, such a model was beyond technological feasibility and data availability given our objectives. It is possible that such models will be built one day by discovering how to model individual people in artificial world laboratories using a bottom-up approach, but this is probably at least 10 to 20 years off. Despite this, it was hoped that all of the most relevant process variables that were available were explicitly included in the model and those that were unavailable were taken into account implicitly by the neural networks that were applied to model agricultural land use. This was probably a most optimistic view, particularly when the forecasts made for 75 years hence assumed a continuation at the same level of all these 'invisible' influences that may or may not have been having a major direct impact. Unfortunately, there was little that could be done about any of this given the urgency of the task and the constraints of technology, knowledge, data, and research resources. In defense of the prototype SPS, there need be no direct relationship between the overall modelling performance and the accuracy of the individual model sub-components and the various data layers. Indeed one of the most compelling justifications for the fuzzy modelling approaches used was the belief that there can come a point in conventional models where improved precision and more detail in the systems of equations can result in a deterioration of performance; see Kosko (1994).

2.3 Major Stages in Operationalising the SPS

Operationalising the prototype system involved:

1. Assembling the available data for a common EU wide spatial framework for the present day.
2. Building a neural net model to provide estimates of population (and any other relevant socio-economic variables) for the same spatial framework.
3. Obtaining and interpolating forecasts of all the variables of interest for about 75 years hence.
4. Applying neural networks to model the relationship between climatic (temperature and rainfall), physical environmental (soil, biomass, elevation) and socio-economic variables to predict contemporary agricultural land use.
5. Forecasting land use for 75 years hence by applying the trained neural networks to classify the forecast values of the inputs under a climatic environmental change scenario.
6. Examining and mapping the forecast changes in agricultural land use.
7. Modifying the forecasts and the predictions to reflect knowledge expressed as fuzzy rules and remapping the impacts in terms of percieved land degradation risk.

3. Assembling the Data

There were major practical difficulties in assembling the database needed for this project. Perhaps the biggest problem was that none of the socio-economic data existed at anything approaching the same geographical scale as the physical and climatic data. Another major problem was that large amounts of GIS pre-processing and often modelling applications were needed to represent all the required data fom multiple different sources in terms of a single consistent spatial framework. Other problems included access to data that existed but were not available for the entire area of interest, cost, copyright and data quality issues. Major differences in quality, scale, and aggregation between the available physical and socio-economic data presented a serious obstacle to straightforward integrated agricultural land use modelling. Physical models of land degradation tended to operate at much more detailed spatial and temporal scale compared with the scale at which socio-economic data were available. Additionally, socio-economic data generally related to historically unstable irregularly shaped zones, whereas physical environmental and climatic models tended to use and produce data in regular gridded structures albeit at a range of different scales. A major reason that much of the existing environmental change research had completely ignored socio-economic systems had already been identified as the lack of socio-economic data at an appropriate level of spatial and temporal detail so as to be directly linked with physical model outputs; see Clark and Rhind (1991). This can still be regarded as a most unfortunate and fundamental obstacle to all research in this very important area.

3.1 Creating a Common Spatial Framework Database

Appropriate available environmental data was acquired and manipulated into the chosen spatial framework - a geographical projection orientated in terms of latitude and longitude on a regular grid at a spatial resolution of 1-decimal-minute (1 DM) using GIS. A grid was selected as the spatial framework in which to store, manipulate, link and map all the data since it offered flexibility in aggregating upwards and could provide a realistic representation of regional or local variation provided the grid cells were sufficiently small in size. A geographical latitude-longitude projection was chosen to compromise 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 DM resolution (approximately 1 km² scale for most of the EU) was guestimated as the most appropriate (smallest practically feasible) level of spatial resolution for this research.

A common spatial framework for all the available climatic, physical and socio-economic data was considered an essential pre-requisite before any integrated modelling could be attempted. Aggregation upwards and interpolation from similar spatial resolution into the chosen spatial framework was relatively trivial compared with interpolation from much coarser levels of spatial resolution. Such disaggregative spatial interpolation was far more problematic and error prone, yet this somewhat essential activity needed to be mastered before much progress could be made. Openshaw and Turner (1999) describes various geo-processing methods developed to handle the problems associated with the aggregation, disaggregation and interpolation of spatial data.

3.2 Digital Elevation Data

The first data set to be processed was the Global Land 1 km Base Elevation source data (GLOBE); see Eidenshink and Faudeen (1996). This 30 arc-second (0.5 DM) resolution grid in a geographical latitude-longitude projection was imported into a proprietary Geographical Information System (GIS) and aggregated to a resolution of 1 DM as follows. Firstly, the 0.5 DM grid was aggregated from each of the four corners of the origin cell of that grid to produce four 1 DM resolution grids where the value in each cell was assigned the mean value of the four 0.5 DM cells from which it was composed. These four 1 DM resolution grids were then disaggregated back to the 0.5 DM resolution where each cell was assigned a quarter of the value at the 1 DM resolution. Now spatial consistent, the four disaggregated grids at a 0.5 DM resolution were combined to produce a single 0.5 DM grid where each cell was assigned the mean of the four incident grid values. Finally, these averaged measurements were aggregated back to a 1 DM resolution (selecting the point of aggregation so as the aggregate data fit the chosen common spatial framework) where again each cell was assigned the mean value of the four 0.5 DM cells from which it was composed. This rather convoluted aggregation procedure reduces the spatial bias in the aggregated data in comparison with a more conventional one step aggregation. In this case, the reason for using the mean is because it is easier to compute than the mode or the median and the resulting 1 DM values approximate to the standard distance units of height above sea level. This variable was required for the whole of the EU as it was used both for the agricultural land use modelling in the Mediterranean climate region and also for the disaggregative spatial interpolation of EUROSTAT socio-economic data.

3.3 Night-time Lights Data

The night-time lights source data is a measure of the frequency of night-time light based on composite satellite images. This (raster) grid of 1 km cells was imported into the GIS, converted into a (vector) polygon coverage and projected from its original Goode Homolosine projection into the required geographical projection using the projection capabilities of the software. The projected polygons were intersected with a polygon coverage coincident with the grid cells of the chosen 1 DM spatial framework. A value was calculated and attached to each intersected polygon by dividing the night-time lights frequency value by the area of the intersected polygon. The intersected polygon values within each 1 DM grid polygon were then added together and the resulting coverage was converted into a grid and clipped to the required size for the socio-economic data interpolation experiments.

3.4 Digital Map Data

Two sources of digital map data were used to create predictor variables for the socio-economic data interpolation exercise: Bartholomew's data (1:1,000,000 scale) and the Digital Chart of the World (1:1,000,000 scale). Data from both sources were combined and manipulated to produce various grids in the chosen spatial framework representing the location of, distance from, or density of either sets of geographical features or values of spatial variables as follows. Firstly all the various map layers were imported into the GIS, re-projected into the geographical latitude-longitude projection and mapped. Geographical features which seemed to be fairly consistently defined across the Mediterranean region of the EU and Great Britain, and thats location or density were believed to be correlated with population density were manipulated into location, cost-distance and density layers. Cells in location layers were bi-valued, depending on whether the cell partially or completely contained either a selected set of geographical features or a selected set of spatial variable values. Distance layer cells were assigned a value corresponding to the latitude-longitude cell distance to the nearest of a selected set of geographical features (or a selected set of spatial variable values). Proprietary GIS provided the simple functionality for point data, yet lacked functionality for creating density layers from line and area data. Density layers can contain more information compared with location and distance layers. For any particular set of geographical features there are many different density surfaces that can be generated and combined, and from these a huge number of location and distance layers can also be generated. Proprietary GIS provided a kernel estimated density routine for point features where it was possible to control the range (bandwidth) of the kernel and specify a weight for the kernel height. For various sets of point features the effect of different kernel bandwidths on the density surfaces was investigated. It was found that as the bandwidth of the kernel increases a larger proportion of the resulting density surface attains non-zero values. In addition, by combining density surfaces made using different kernel bandwidths even more useful density information could be created, and in each case the most useful cross-scale combination was dependent on the distribution of the geographical point features. After extensive mapping and analysis it was discovered that combining kernel density surfaces for overlapping cross-scale ranges of kernel widths by adding their values scaled relative to the size of the kernel which produced them offered the most useful density information for the impending disaggregative interpolation experiments. So, in this way cross-scale density surfaces were created for a various point features including railway stations that were expected to be positively correlated with population density and mountain summits that were expected to be negatively correlated with population density. A proprietary GIS macro program was written to generate similar cross-scale kernel density information for line and area data.

3.5 Climate Data

Climate data were supplied by the MEDALUS III team from the Climatology Research Unit (CRU) at the University of East Anglia. The base-line data was produced by interpolating measurements from a network of about 50 weather stations across the northern Mediterranean to produce 0.5 DM resolution seasonal average temperature and precipitation totals, see Palutikof and Agnew (1997) for an explanation of the statistical down scaling procedure. The base-line data was imported into the proprietary GIS and the nested cells were aggregated to produce the desired 1 DM resolution data. For temperature data the mean of the smaller grid values were used and for the precipitation data the sum of the smaller grid values were used. In addition to the base-line data for 1970-79, forecasts of future seasonal average temperature and precipitation totals for 2030-39 and 2070-79 based on global climate change models were also made available by the CRU. These data were linearly interpolated to produce maps of temperature and precipitation for around about now and 75 years hence. The levels of spatial uncertainty in these transformed climatic data were matched by equal or greater amounts of climatic scenario forecast uncertainty.

3.6 Soil Data

From the Soils geographical database of Europe (at scale 1:1,000,000) the 26 soil types were classed into six major groups with the help of a soil expert and converted into location grids. Soil quality measures were then developed to make use of the information about the characteristics of the soil in the source data. The fundamental physical properties of soil profiles (soil type, rooting depth, soil texture, water regime, slope, and the existence of impermeable layers) were combined to produce soil quality layers by coding the expert knowledge of the soil scientist into a set of fuzzy rules using a mathematical software package with fuzzy inference capabilities. The soil quality layers were developed in this way so as to make use of the information regarding the characteristics of soils in the soil database without having to add each as a separate input into the model. These soil quality layers were developed in a similar way to a general land use capability classification, but in a simplified way that did not take into account all the physical interactions between the soil and climate; see Burrough  et al. (1992).

3.7 Biomass Data

Estimates of biomass were provided by MEDALUS III colleagues researching at the University of Leeds. These were outputs from a model which translated temperature and rainfall data via a set of equations and proprietary GIS operations into estimates of potential biomass; see Kirkby et al.  (1997); Harris (1998). The climatic biomass model was based purely on climate data and did not take into account factors like the height above sea level or soil type. The data became available at a 0.5 DM level of resolution towards the end of the project, (initially it had only been available at a relatively course 30 DM resolution which needed interpolating into the desired spatial framework). The 0.5 DM resolution output was aggregated in the same way as the temperature data to produce estimates for the chosen common 1 DM spatial framework and was used in the modelling reported here.

3.8 Land Use Data

Several different land use data sources were used including: the soils geographical database version 3.2 which contained a classification of dominant and secondary land use; the US Geological Survey (USGS) land use and land-cover classifications, including the Seasonal Land Cover Regions (SLCR) classification, see Anderson et al. (1976); the International Global Biosphere Programme (IGBP) classification, see Belward (1996); the Biosphere-Atmosphere Transfer Scheme (BATS) classification, see Dickinson et al. (1986); the Global Ecosystems (GE) classifications, see Olson (1994a, 1994b); and the Simple Biosphere Model (SBM), see Sellers et al. (1996). Multiple land use classification data sets were used because there is both considerable uncertainty regarding the classification of land use and also variation in the details obtained from satellite imagery; for example, the SLCR had 255 different land use classes whereas the SBM had only 10. The land use modelling could have been greatly improved by using the probabilities of each cell belonging to particular land use classes based on the original satellite data (before it was classified); see Moody et al (1996) and Carpenter et al (1997), however, these data were unavailable for this research. For each of the different land use classifications the classes were aggregated into four groups; arable or crop, tree or orchard, shrub, and waste or barren. Each aggregated group was then converted into a location grid. In tune with the fuzzy approach, if one of the initial classes could belong to more than one aggregate group it belonged to both, so the aggregate land use classes were not necessarily crisp.

In addition to the aforementioned land use data, normalised difference vegetation index (NDVI) land cover data were obtained from the USGS. The NDVI is an indicator of greenness or vegetation abundance derived from AVHRR satellite imagery that was to generate additional land degradation risk indicators. These data are available at a spatial resolution of about 1 km² and were obtained along with much of the land use data from the following URL where a full description of the data is provided:
http://edcwww.cr.usgs.gov/landdaac/glcc/glcc.html

The source NDVI data were obtained in a Lambert-Azimuthal equal area projection and reprojected into the chosen latitude-longitude projection and subjected to aggregation using the methods previously described so as to fit the chosen spatial framework. These data are monthly composites that were aggregated into seasonal values and clipped to the area of interest.

3.9 Population Data

Demographic data at a fine level of geographical resolution were only available for the Great Britain and were estimated using a neural net based interpolation procedure for the rest of the EU. The source data used to train the neural nets were based on the 1991 grid-square population estimates for the UK known as Surpop. The Surpop 0.2 km resolution source data were manipulated into target data for the socio-economic data interpolation experiments in the following way. Firstly, the grid was imported into the proprietary GIS and converted into a polygon coverage in its source Ordnance Survey projection. The polygon coverage was then projected into the geographical latitude-longitude projection of the common spatial framework. These projected polygons were then intersected with a polygon coverage that coincides with a 0.125 DM grid that neatly aggregates to the chosen 1 DM resolution spatial framework. 0.125 DM sized polygons were used since they were approximately the same size as the projected polygons. The intersected polygons were assigned proportions of the population depending on the area of the intersection, then the intersected polygon values within each 0.125 DM polygon were summed and the resulting coverage was converted into a 0.125 DM 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 same aggregation disaggregation re-aggregation procedure described previously was then used iteratively until the population values fit neatly into the chosen spatial framework.

NUTS3 and NUTS2 resolution population data from EUROSTAT, NUTS2 population forecasts from the Netherlands Interdisciplinary Demographic Institute (NIDI), 1991 UK Small Area Statistics census data, Italian National Statistical Office (ISTAT) population data for registration zone centroids, the World City Population Database (WCPD), and RIVM and Tobler's pychnophilactic population density estimates for Europe were also used in the socio-economic interpolation experiments. These data were manipulated into the common spatial framework using the techniques described above prior to being used in the modelling described below.

3.10 Socio-economic Data Modelling

So, the initial modelling task was to use the aggregated Surpop population surface to train a neural net based spatial interpolator to relate population density to a selected set of predictor variables. For each 1 DM cell in the spatial framework, the values of variables chosen to model the population density were concatenated into a large file of vectors from which randomly selected training data sets of 10,000 were created. Training and validation data for EU Mediterranean regions would have been used if they were made, but unfortunately none were. As a consequence, detailed socio-economic and demographic breakdowns were too error prone to be worthwhile. A major source of uncertainty still concerns how different the population density relationships with the predictor variables in this region were compared with those in Great Britain. We felt that the EU needed to organise its basic data resources to a far better degree to tackle this problem. It seemed most unsatisfactory that even NUTS 5 (English ward level) resolution data were not available throughout the EU and that data copyright and ownership prevented access to existing high resolution data for this research.

4. Agricultural Land Use Modelling

Land use change and land degradation are complex interacting geographical processes which are themselves intricately related to climatic, physical and socio-economic change. Land use reflects land-capability that depends both on climatic interactions with other physical characteristics of the environment, and also land suitability that is complexly interrelated with various socio-economic factors including the characteristics of market supply and demand.

The modelling methodology can be summarised in the following seven steps:

4.1 Predicting Contemporary Land Use and Land-cover

Step 2 was to train neural nets to model agricultural land use based on patterns between; soil type, soil quality, potential biomass, average air temperature (in spring, summer, autumn and winter), average monthly precipitation (in spring, summer, autumn and winter), height above sea level, and population density. A list of predictors is given in Table 1. Different neural nets were trained to predict arable or crop, tree or orchard, shrub, and waste or barren agricultural land use groups. Recall that various different aggregated land use classifications were used to cope with the lack of unclassified remote sensing data that should have enabled the predictions to be based on probabilities. Anyway, the neural net predictions were used to create maps of predicted agricultural land use where the fit seems remarkably good; see for example, Figure 1 showing predicted waste or barren land and the observed for an aggregate classification.

NDVI land-cover data modelling was undertaken in two stages: (1) predicting and forecasting the NDVI data based purely on physical and climatic variables and (2) predicting and forecasting also using socio-economic data inputs. The results provided some evidence of socio-economic effects as errors significantly reduced when the socio-economic inputs were included. The results were used further to generate a land degradation risk indicator that was combined with the land use change based indicator in subsequent fuzzy inference to demonstrate how to form a scenario based consensus, (they might also have been input as indicators in the land use modelling but were not).

4.2 Forecasting Land Use and Land-cover

5. Translating Agricultural Land Use Changes into Land Degradation Terms

The next major step involved translating agricultural land use changes into land degradation terms. The way in which environmental and economic interpretations of agricultural land degradation differ is important. Environmental interpretations tend to be geared towards things like soil quality, biodiversity, biomass, habitat preservation and conservation erring somewhat away from the long term exploitation of natural resources, whereas economic interpretations tend to give greater emphasis to things like crop revenue and land prices erring more towards short term exploitation of natural resources. Consider for example an area of land where only the aggregated tree or orchard land use class has been predicted as a contemporary land use and only the aggregated crop land use has been forecast as a future land use. A more environmental perspective might regard land degradation risk to be high because of likely reductions in biodiversity and biomass and increased soil erosion risk. A more economic viewpoint regarding likely increases in income from crops and potential economic multiplier effects from job creation may contend that the land use change is a land improvement and not a degradation.

Mathematical software capable of fuzzy inference was used to translate the land use changes into land degradation risk surfaces and further combine these with other environmental risk indicators shown in Table 2. An illustrative set of the fuzzy rules are given in Table 3, and the resulting synoptic scenario based forecast consensus map of land degradation is shown in Figure  3. A superficial interpretation might be that Italy is relatively less prone to land-degradation than the Iberian peninsular and that Greece is perhaps intermediate under the given climatic change scenario.

The map shown in Figure 3 can be interpreted at a synoptic scale and it is conceivable that such maps could be used by land use planners as decision making aids for allocating funds to implement land improvement schemes. Maps are essentially decision making aids and the prototype SPS could be developed to aid spatial decision support relating to land degradation by helping formulate subsidy policy to control agricultural land use change. With any luck the prototype will encourage informed debate and in the end may provide a means for politicians to justify fund allocation at national, regional and possibly even local scales without the need for a deep understanding of the science involved. The map of forecast land degradation for 75 years hence shown in Figure 3 is based on our personal subjective interpretation of land use change related land degradation risk and its perceived importance relative to other land degradation risk indicators as translated via the set of fuzzy rules given in Table 3. Given a number of such subjective opinions from a range of experts further consensus can be made to reduce the level of subjectivity. To facilitate this we hope to develop a web- front-end to the SPS and make the analysis interactive.

6. Conclusions

This paper has outlined a prototype Synoptic Prediction System that provides broad brush agricultural land use and land degradation risk forecasts for the 21^st century. The results reflect the available data and our own subjective opinion of the land degradation concept. The prototype captures the very essence of a GIS based approach to modelling environmental systems; it is broad-brush, it is visual, it is synoptic, and it is broadly representative. The main aim of this research was to raise awareness of the possible impacts of climate change on land use in the next century and suggest tools and approaches that can cope with the uncertainties involved in forecasting land use change. The research challenge remains and the modelling can be greatly improved with further research efforts as better data and process knowledge becomes available. The results only serve as a preliminary test of methodology, they are too subjective still to be regarded as a legitimate benchmark. In our opinion, the prototype system appears to provide a useful framework for assessing the possible impacts of climatic change on agricultural land use by linking all the various components in a novel and interesting way. The prototype simultaneously demonstrates what is needed to model the process of land degradation as well as giving some reasoned indication of the likely effects of climate change on Mediterranean agricultural land use. If these results are really what is required then no doubt their accuracy can be improved by further research.

Further details of the research can be found on the internet at the following URL:
http://www.ccg.leeds.ac.uk/projects/medalus3

References

Tables and figures

Figure 1. Observed and predicted waste or barren land areas

Figure 2. Forecast waste or barren land areas

Figure 3. Land degradation forecast

Table 1. Predictors used to model and forecast agricultural land use

Table 2. Variables used in the fuzzy translation of land use into land degradation terms

Table 3. Example fuzzy rules used to generate a land degradation (LD) output map

Data Sources

1. Bartholomew's European 1 DM resolution digital map data set.
2. The Digital Chart of the World (DCW).
3. The Defence Meteorological Satellite Program (DMSP) Operational Linescan System (OLS) Night-time lights frequency dataset.
4. RegioMap (EUROSTAT) CD-ROM socio-economic data for EU NUTS regions.
5. Tobler's pycnophylactic (mass preserving) smooth interpolated population density surface.
6. RIVM's raw population count for geographical regions.
7. UK Census data: Surpop 200 meter total population and population seeking work surfaces of Great Britain; SAS Small Area Statistics.
8. Italian National Statistical Institute Registration total population count point data.
9. MEDALUS III temperature and precipitation data from CRU at UEA.
10. Climatic Biomass Potential data from MEDALUS III at Leeds.
11. Soils geographical database of Europe at scale 1:1,000,000 version 3.2.
12. GLOBE: Global Land One-KM Base Elevation Data version 0.1.
13. United States Geological Survey (USGS) land-use and land-cover classifications including: Normalised Difference Vegetation Index (NDVI), International Global Biosphere Programme (IGBP), Seasonal Land Cover Regions (SLCR), Biosphere-Atmosphere Transfer Scheme (BATS), Simple Biosphere Model (SiB).