Modelling climate envelopes for rare Madagascan Butterflies
	Modelling climate change effects on African Plant species
	>Introduction
	Genetic Algorithms
	Datasets Aggregation
	Methodology
	Continental Scale Modelling
	Future Work
	Workshop
	References
	Correspondence

This work is funded by:

Introduction

Citation: Tokumine, S. 2002. Bioclimatic modelling of sub-Saharan African plants. Progress report to Conservation International, September 2002. Centre for Ecology Law and Policy, Environment Department, University of York, UK.

Global climate is changing, and is projected to continue changing (Hughes, 2000). At large scales of 10 km2 and above, macroclimate has been seen as a crucial element in the distribution patterns of many organisms (Huntley et al. 1995). The rate of genetic adaptation is unlikely to match the speed of climate change (Etterson & Shaw 2001). This change is already believed to have had an impact on many natural systems (IPCC 2001a), or is predicted to cause major changes to biodiversity for which new conservation paradigms must be established (Peterson et al. 2002; Hannah et al. 2002).

These new paradigms need predictions of potential future change on which to base current conservation strategy. Ideally, this information would be provided by long-term, controlled lab experiments monitoring the direct effects of changing climatic variables (e.g. Davis et al. 1998). Ideally, these lab experiments would then be combined with extensive field experiments in all the major vegetation types of the world, so nurturing a deep understanding of the processes underlying the community structure in all cases (Johnston & Schmitz, 1997). It's unfortunate that this knowledge is likely to remain unavailable in the timeframe for dramatic climate change predicted by the IPCC (Root & Schneider, 1993; IPCC, 2001a).

One commonly used alternative approach to predict the effect on the distribution of biota under climate change is based on the assumption that species distributions are directly dependant on local climate (Walter, 1979). At its simplest it involves "linking" species' current distributions with combinations of current climate data, then plotting spatial shifts of these "climate envelopes" using climate change scenario data. Global geological histories provide clues as to the possible effects of climate change on the biota of the earth (Graham & Grimm, 1990; Webb, 1987). Changes in the distribution patterns of these terrestrial organisms have been shown to be the primary response to paleological climate changes with macroevolutionary response to climatic fluctuations appearing limited in comparison (Huntley & Webb, 1989). Though we cannot be certain past changes are directly comparable to the changes we will experience due to the differing combination of causal factors (Graham & Grimm, 1990), we can expect the distributions of current plant species to exhibit a distributional response with future changes in global climate. The linking approach has come under criticism for not incorporating the possible impacts of changing inter and intra-species interactions under different climate change scenarios (Davis et al., 1998). Though acquisition of this knowledge is important, its absence is symptomatic of the lack of experimentation highlighted above. "Linking" techniques are therefore contemporarily important as 'null models' from which we can view the possible changes due to global warming.

Back to top

Previous methods of implementing the linking approach are broadly based on two methods, generalised linear models and BIOCLIM (Nix, 1986) approaches. Generalised linear models have been mainly used in instances where there are largely complete datasets incorporating absence data (Yee & Mitchell 1991). An extension of this has seen the use of climate response surfaces (Bartlein et al. 1986; Hill, Thomas & Huntley, 1999). BIOCLIM-type approaches have been used on less complete datasets which focus on species presences (Busby, 1986; Nix, 1986, Eeley et al.1999). Combinations of the two approaches have also been proposed under the Genetic Algorithm for Rule Set Prediction (GARP) project, which uses a combination of BIOCLIM rules, logistic regression and machine learning methods. (Peterson et al. 2002).

BIOCLIM type approaches are based around evaluating a species' "Climate Envelope". These are areas defined by the overlay of a number of ranges of climate variables. These ranges describe the minimum and maximum values of a climate variable found at the location where a species occurrence is recorded. In this way, all areas exhibiting a combination of climatic conditions within the range of conditions dictated by a species' distribution are found. This method therefore delineates climatically suitable areas for the species, in other words, their "climate envelope". If evaluating a species using two climate variables, annual rainfall and temperature for example, we may find the range of rainfall and temperature experienced across all points in the species distribution to be between 150-250 mm and 10-15oC respectively. All areas within the study area that satisfy both these requirements would make up this species' climate envelope. As more climate variables are added, the description of suitable climate becomes increasingly specific to the species distribution, resulting in a climate envelope more spatially representative of that species distribution. However, there are limits to the extent to which this process is practicable. Using typically available sets of relevant climate variables, the overlay technique of BIOCLIM can often result in over-predictions of suitable areas.

Removal of a point in a given species distribution can result in a more narrowly defined climate envelope, potentially leading to a reduction in these overpredictions. Minimising overpredictions while maximising the climate envelopes capture of actual species occurrences becomes a balancing act. A number of indices exist that quantify this trade-off. These indices are functions of the number of correctly and over or under predicted areas, addressing issues of reliability important in many areas of GIS and remote sensing (Story & Congalton, 1986). Though the Kappa statistic has been suggested as an index of fit which most reliably represents the fit of a climate envelope to a given species original distribution (Manel et al., 2001), the index requires absence data often lacking from library species distribution records. In order to assess the quality of a climate envelope in spatially delineating a given species' distribution where only presence data is available, a normalising similarity indices such as Sørenson's Index can be used (Clifford & Stephenson, 1975).

Back to top

To find the climate envelope for a given set of climate variables with the best possible fit to the actual species distribution according to Sørenson's Index , we could construct a climate envelope and calculate the Sørenson's Index for all possible combinations of points of a species distribution. For small ranging species, all possible combinations can easily be computed. However, to assess the climate envelopes for all combinations of a species distribution comprising, for example, 50 points would result in a number of possible combinations so large the feasibility of this "brute force" approach is limited.
The problem of finding this optimal index score in a field of possible solutions is a classic example of a "NP - complete" Boolean Satisfiability Problem (SAT); a class of computational problems for which no efficient solution algorithm has been found (DeJong & Spears, 1989). When an NP - complete problem must be solved, one approach is to use computational heuristics to obtain an optimal or near-optimal solution. NP - complete SAT's have been shown to be effectively solved by genetic algorithms (GAs) (DeJong & Spears, 1989). GAs are adaptive heuristic search algorithms based on the idea of natural selection. Initially developed by Holland (1975), they have been widely applied in many fields where there are NP-complete problems from fields as diverse as music and circuit board design. GAs have also been used in spatial analysis. Applications range from the definition of catchment areas for building society branches (Hobbs 1995) to optimal patch configuration (Brookes, 2001). One approach to defining climate envelopes using a GA to develop decision rules is also reported by Stockwell and Peters (1999).

In addition to providing a method to find the near optimal climate envelope, the heuristic optimisation approach of GAs has certain advantages over more traditional statistical approaches to creating a predictive model of species presence/absence. Notably, possible regression techniques such as logistic regression may be affected by over-dispersion caused by model miss-specification. This can result from the spatial autocorrelation found in the climatic independent variables and the dependent variable itself. It also may make intuitive ecological sense to consider ranges of climate variable values that may be suitable for the occurrence of a species, rather than using statistical approaches implying single, optimal, variable values associated with species area occurrence.
Once defined, by substituting the current, observed environmental variables used to define the climatic envelope with those derived from future atmospheric general circulation models such as HadCM3 (Gordon et al. 2000), potential spatial changes in a species distribution are elucidated. Standard GIS software is used for the manipulation of the present and future climate surfaces, along with the species spatial distributions considered. Once analysed by the GA, results from the consideration of these surfaces are returned to the GIS for mapping and the calculation of overlap between observed and predicted species distributions. The GIS also allows an assessment of the spatial concurrence of present observed species distributions and future predictions under climate change.

We describe a tested, GA based methodology for the modelling of species distributions in climate space for which only presence data is available. We also describe the development of Tanzanian Eastern Arc tree species datasets, Tanzanian DEM aggregation results, present and future climate and soil surface datasets at the continental scale and emergent issues from this. These components are discussed within the context of high resolution Tanzania scale modelling of 452 tree taxa and low resolution continental scale climate envelope modelling of over 3500 plant species.

Key Achievements:

1. African continent scale climate change surfaces for the present day, 2025, 2055, 2085.

2.High resolution Tanzanian Digital Elevation Models (DEM) and soil datasets.

3.High resolution Tanzanian Tree species dataset consisting of 452 tree taxa at 220 locations across the Eastern Arc Mountains.

4.Genetic algorithm software program development complete.

5.Linking and automation program between the GA and WORLDMAP plant mapping software development complete.

6.3566 individual genetic algorithm optimised climate envelopes constructed for the entire sub-Saharan Africa continental WORLDMAP plant database (~10% of total African flora). Allowing accurate reproduction of species richness and endemism using only climate variables (Figure 1 & 2).

7.Projection of all 3566 climate envelopes to climate predicted for 2025, 2055 & 2085, together with per species information on extinction, colonisation, and species turnover (table 4).

8.Genetic Algorithm methodology submitted to the journal of Computers Environment and Urban Systems.

9.Genetic Algorithm methodology presented at the GIS researchers UK conference, the British Ecological Society macroecology conference, and at a workshop hosted by this project at the University of York.

Continue|Back to top