Biodiversity and Conservation 9: 869–885, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
A simple non-parametric GIS model for predicting species distribution: endemic birds in Bioko Island, West Africa STUART M. LENTON1 , JOHN E. FA2,∗ and JAIME PEREZ DEL VAL3 1 R & D, Heath International Non-Marine, 133 Houndsditch, London EC3A 7AH; 2 Durrell Wildlife Conservation Trust, Les Augrès Manor, Jersey JE3 5BP, UK; 3 Museo de Ciencias Naturales, CSIC,
c/ José Gutierrez Abascal 2, 28006 Madrid, Spain; *Author for correspondence (fax: +44 1534 860002; e-mail:
[email protected]) Received 8 April 1999; accepted in revised form 16 November 1999
Abstract. Species mapping is a useful conservation tool for predicting patterns of biological diversity, or identifying geographical areas of conservation significance. Mapping can also improve our understanding of the appropriateness of habitat areas for individual species. We outline a computer-based methodology, PREDICT, for the analysis of the habitat requirements of species in a combined GIS-statistical programming environment. The paper details the statistical background to the approach adopted, the program structure and input file information and then applies these techniques to bird data from Bioko Island, West Africa. It produces images and statistics that assess the potential of unstudied areas for wildlife for which presence/absence data and basic habitat information are available. Suitability for target species is determined within surveyed and non-surveyed squares by a form of weights of evidence. The program measures the degree of association between habitat factors and presence/absence of target species by means of χ 2 tests. The overall suitability weighting of each square, as the sum of all individual habitat factor weightings, is finally displayed in maps depicting areas of highly suitable, suitable, unsuitable and highly unsuitable habitat. Statistical relations between vegetation, rainfall and landscape features on Bioko Island and the location of 9 endemic bird taxa are presented herein. Final confirmation of the accuracy of predictions of the studied bird taxa will ensue from future field observations. However, in a series of misclassification tests of the program, actual distribution detection rate was in excess of 90%. The use of PREDICT can guide investigations of little known species in remote areas and provide a practical solution to identify areas of high rare species diversity in need of conservation. Key words: Bioko Island, endemic birds, GIS, habitat suitability
Introduction Species distribution mapping is of fundamental importance to our understanding of biodiversity patterns and as an applied tool for conservation managers (Bailey and Hogg 1986; Fa and Morales 1993; Miller 1994). The simplest distribution maps, and probably the most objective but least informative, are those that are created by drawing boundaries around the location points where specimens have been collected. These maps have a limited use in habitat–wildlife relationship studies since inconsistencies in coverage can lead to erroneous conclusions (Scott et al. 1993). There
870 is a large volume of literature on methods for modelling wildlife distribution including habitat suitability indices (HSI), statistical methods including generalised linear models (GLMs) and generalised additive modelling, as well as spatial and inductive (cartographic and regression trees, neural nets, Bayesian and weights of evidence) models (see Walker 1990; Butterfield et al. 1994; Miller et al. 1989). For example, Osborne and Tigar (1992) used GLM for predicting probabilities of bird occurrence in Lesotho. HSI models test the ability of habitat to provide life requisites for a particular species, by determining a numerical index of appropriateness and assigning that value to a given land-use category or vegetation type at a given point in time (Lancia et al. 1986). Such models are useful for species management planning as well as for impact-assessment studies (Fish and Wildlife Service 1981). Such models can be applied at little cost and produce outputs that are easily understood. Morrison et al. (1992) provide examples of studies that link ecological processes and GIS (geographic information system). The latter studies employ packages such as RAMAS/GIS (Akçakaya 1996) which link GIS and metapopulation models which assess habitat suitability for individuals of species, not populations. While these models are robust, their dependence on parametric statistics restrict their applicability to databases of unknown statistical distributions, and of small number of observations. There is a need to develop models that apply a modern, spatially explicit approach to predicting wildlife distribution based upon habitat characteristics. We present a model which takes into account the practical realities of poor data availability and quality in many parts of the world and which can be run on portable computers with modest storage and processing. This goes against the current trend of applying ever more sophisticated computing and mathematical techniques to predicting complex spatial distributions. Even in well-mapped countries source data of sufficient quality to match these approaches may be difficult to obtain, whilst for most of the world it is often impossible. Our main objective was to develop a cost-effective habitat suitability GIS model that would involve using non-parametric statistics (Unwin 1995). We also wanted to create a technique that could produce maps of sufficient resolution, in short run-times, to assist conservation biologists delimit and refine areas of possible occurrence of a species whilst in the field. We aimed to produce a module applicable to any species held within a database, without requiring amendments to the program or to the habitat datasets. Such modules would be particularly valuable in little known areas, where considerable investment in time and money would otherwise be required if only ground surveys had to be undertaken. We describe a small-sized statistical program (61 kb), PREDICT, written in C for a DOS platform, which can operate stand-alone or be incorporated into current GIS. The model employs minimal coverage of surveyed squares to assess habitat suitability within a larger study area. It uses the correlations of presence/absence data of the species with habitat factors contained in a geographical database.
871 We use presence/absence data of the endemic birds of Bioko Island, West Africa (43 subspecies and 2 species out of the 143 resident taxa), to illustrate the model’s output. Although output data are presented herein, only maps of the 2 endemic species and of 7 subspecies are shown. These taxa were selected because they represent a range of contrasting distributions amongst the endemic island birds, independent of output map quality. To investigate the predictive capability of the model we carried out a number of misclassification tests using the current datasets. Use of the model for prioritising conservation areas for the endemic Bioko birds is presented elsewhere (Fa et al., in preparation). Study area Bioko is a continental island (69 × 32 km; 2017 km2 ), 32 km from the Cameroon coast. Two volcanic massifs dominate the island, the Caldera de Luba (2261 m) in the southwest, and Pico Basilé (3011 m) in the north. Bioko’s vegetation occurs in altitudinal belts (Fa 1992). Rainforest formations are typical between sea level and 800 m (most of this been cleared and transformed into cacao agrosystems or other food plantations), montane forest appearing between 800 and 1800 m. This is succeeded by mossy forest at 1800–2500 m. Above 2400 m, heaths (Ericaceae) and grasslands are found. Climate is tropical equatorial, influenced by the north–south movements of the Intertropical Convergence Zone. Weather is affected by the island’s rugged landscape; the southern portion of the island receiving over 10,000 mm of rain annually, whereas the north has just 2000 mm. Methods Environmental variables Environmental information consisted of: (a) landform/altitude data derived from the Spanish Instituto Geografico Nacional’s 1:100,000 topographic map (IGN 1979); slope angle, slope aspect and elevation obtained from a digital elevation model (DEM) of the topographic map; (b) vegetation type taken from a distribution map of the main vegetation categories (FED/DHV 1990, sensu Ocaña Garcia 1961; Juste and Fa, 1994; Juste 1992); and (c) rainfall data from Perez del Val’s (1994) isohyet map. Bird distribution Presence/absence data of Bioko birds were obtained from census counts (mist-netting and direct observations) recorded by Perez del Val (1994) within 31 sample points
872 throughout the island. Bird censuses were conducted within all habitat types. However, because of inaccessibility, some areas were not covered. Data capture, transformation and editing GIS ARC/INFO (Environmental Systems Research Institute, Inc., Redlands, California; Environ. Sys. Res. Inst. 1989) was used to convert all maps into digital format before transferring to IDRISI, a PC-based raster GIS (The Idrisi Project, Clark University Graduate School of Geography, Worcester, MA 01610-1477). IDRISI was chosen because it can incorporate modules written in most programming languages, its cost-effectiveness and the possibility of using it directly on a portable PC under field conditions. All cartographic datasets were converted to a common co-ordinate system. Affine transformations, to correct effects of scale, shift and rotation, were applied to the topographic datasets, using grid intersections (n = 41) as control points. The root mean square error (RMSE) for the transformation was 0.2 mm (at the plotted scale of the source document). Separately, eight strategic points were selected around the coastline or on clearly identifiable kinks and pinnacles for the rainfall and vegetation maps. This was necessary because these maps were not plotted to a fixed scale with no grid points to reference the scanned image. RMSE for the rainfall map was ±0.6 mm (scale 1:480,000) and ±1.0 mm for the vegetation map (scale 1:360,000). Data editing on some 5635 arcs produced by the raster-to-vector conversion was undertaken in AUTOCAD (Autodesk Ltd, Cross Lanes, Guildford, Surrey, UK). Model choice GIS models can be prescriptive or predictive (Bonham-Carter 1994). Prescriptive models involve overlaying binary maps using Boolean procedures. This technique denotes suitable (1) or unsuitable areas (0), with no middle ground. For wildlife, such maps can be created for a species provided its habitat preferences are well researched and rules can be specified accordingly. Being essentially deterministic, such rules are only applicable for the chosen species. In contrast, predictive models use less strictly defined criteria and can be described probabilistically. These models are data-driven (based on observations) or knowledge-driven (relying on input from experts). Knowledge-driven models employ Bayesian probability and fuzzy logic. If sufficient data is available, log-linear Bayesian methods can estimate the importance of evidence from the data (Bonham-Carter 1994). However, results from these models depend on data quality; anyone other than an expert in the field risks producing misleading or inaccurate results. Additionally, data on a particular species may be limited, or few people may have the profound knowledge needed to operate such a technique. In contrast, data-driven methods use logistic regression or weights of evidence. Because logistic regression is not a viable option for studies with only small numbers of
873 observations, as for rare species, a form of weights of evidence is applicable through the construction of contingency tables. Statistical procedures We adopted a form of indexed overlay for our model (Bonham-Carter 1994). For each species, the significance of each habitat factor for the species must be known before overlaying. Once the significance of the factor is known, each category within that factor must also be scored in some way. If altitude is a significant factor to the location of a species it follows that the chances of finding the species at all altitudes are not the same as when altitude is insignificant. The resultant weighting for each habitat factor must then be applied across the whole area, according to the individual categories found at each location. Given the nature of the Bioko dataset, all habitat factors were classified into four categories (Figure 1). For each species the model uses a twoway contingency table for each habitat factor alongside presence/absence data at each observation site. Such tables use cross-classified nominal or ordinal categorical data (O’Brien 1989). For the Bioko bird data, the resulting matrix is 2 × 4 as shown in the example below: Habitat category 1
2
3
4
Total
Present Absent
0 5
4 7
4 4
7 0
15 16
Total
5
11
8
7
31
χ 2 tests can then be employed to assess whether row and column variables of the contingency table are independent of each other. The null hypothesis is set up in which the two datasets (i.e. species observations and a certain habitat factor) are assumed to be independent. If the null hypothesis can be rejected, the two datasets are indeed associated in some way. Alternatively, if there is insufficient evidence to reject the null hypothesis, the conclusion must be that the two datasets are independent, and that knowledge of a value in one dataset gives no indication of its corresponding value in the other dataset. Independence is assessed by calculating the expected frequencies given the null hypothesis, and by comparing these with the observed frequencies to ascertain whether or not they are significantly different. The observed data can be considered a sample and as such is subject to sampling error. Given this fact, it is statistically possible that the observed dataset of bird locations is the product of a random, or chance, variation and therefore not a true reflection of any relationship between the species and habitat factor in question. To determine the likelihood of this being the case an overall χ 2 statistic can be computed and assessed against a theoretical benchmark, namely the χ 2 distribution.
.
Figure 1. Habitat factor maps used in the study of bird endemics in Bioko Island.
874
875 The modelling procedure Our program is designed to run in conjunction with IDRISI and produces images that can be viewed directly without further processing. It can, however, be used with other systems, although some additional manipulation may be required to display the results. The program operates on image files in their simplest format. In IDRISI, these consist of a single column of figures. In addition to this, there is a documentation file, which provides the header information required for display including details such as map extent, number of rows and columns and pixel resolution. The prediction program reads in values for successive pixels from each of the habitat datasets, performs analysis based upon values retrieved from the database, and finally produces a suitability image in the same format. The full algorithm is summarised in Figure 2. The suitability image is produced in two stages: (a) the creation of an empty image file with the same pixel resolution and extents as the input maps, and (b) the application of the weights for each habitat factor in turn to that image. This process can be viewed as a constant stream of pixels, which receive updates to their value depending on the significance of the specific habitat type found at that location for each habitat factor. Positive and negative values indicating suitable and unsuitable areas respectively are then reclassified into bands representing: 1. highly unsuitable (< −1.0); 2. unsuitable (≥ −1.0 and 0.0 and ≤1.0) and 4. highly suitable (>1.0). Zero values cannot be identified as either suitable or unsuitable. In the unlikely event that a pixel did receive a suitability value of 0, the assumption is made that there was insufficient evidence to make a prediction. The final stages of the program are devoted to carrying out checks on the suitability image created. Results of these checks are reported to the user via a series of text files. These files include: (a) details of the suitability value generated for each of the observation sites (which can be directly compared with abundance figures not used by the program if available) and (b) a summary of the number of pixels assigned to each suitability band thereby allowing the user to quantify the extent of suitable and unsuitable areas, or compare between different species or scenarios and (c) a detailed breakdown of each suitability band providing pixel counts for each category in each dataset.
Results Habitat suitability maps Light grey and grey pixels in the generated maps for the 2 endemic species and the 7 endemic subspecies, represent unsuitable areas where it is extremely unlikely that the species in question would be found (Figure 3). Dark grey and black zones, in
876 contrast, indicate areas that are suitable, and highly suitable, respectively; the probability of finding the particular species in any of these areas is thus relatively high. However, it should be remembered that this is by no means a guarantee that you will find the particular species, but merely an indicationthat the combination of those
Figure 2. Algorithm for modelling habitat suitability of wildlife using PREDICT.
877 habitat factors assessed for that location provides a suitable environment. As a check to the modelling procedure the suitability values generated for each of the observation sites were closely monitored. Results proved that, as expected, none of the sites where a species was found to be present received negative values. The pixel effect on the images is created by the 500 m cell resolution used in this project, but can be reduced by using a smaller resolution. However, this will substantially increase the size of files and also increase the execution time of the program. For the purposes of this project, a smaller resolution is not considered necessary given the size and mobility of the study subjects. The program itself is not affected by cell resolution except in the time that it takes to run. Significance of habitat factors on bird distribution The degree of association (in terms of the χ 2 probabilities) between bird observations and habitat factors was typical for vegetation, altitude, rainfall and aspect (Table 1). Average probability values for slope were low in comparison. Misclassification testing of the model Misclassification testing is complex for the type of model developed. The most appropriate test would be to collect new data in a sampling regime stratified by the current model. For the present purpose, we carried out a misclassification routine using data from points selected within delineated altitude and vegetation bands (Figure 4). From these, a database of presence/absence was derived for the same 31 observation sites used to collect the bird data. PREDICT was used to re-create the ‘known’ distributions using only information stored in the database. Analysis of the results showed that, in the majority of cases, over 90% of the distribution (93.67 ± 6.2%; median: 97.0%) was detected. Test distributions and derived observations are shown in Figure 4 along with their predicted images.
Discussion The analytical functionality of GIS is already providing a major component in many conservation projects and their use can only become more widespread in the future, as tailor-made solutions become available. However, conservation studies are frequently forced to operate with strictly limited resources, both in terms of time and finances. PREDICT provides a cheap and easy method that can be used to guide the sampling of little-studied species even in remote areas. The model is an easily used practical application employed here for the identification of areas of importance for the conservation of the endemic birds in Bioko. A further value of the model is its ability
878 Speirops brunneus
Ploceus albinucha maxwelli
Tersiphone rufiventer tricolor
Estrilda nonnula elizae
Pogoniulus scolopaceus stellatus
Batis poensis
Zosterops senegalensis poensis
Lamprotornis splendidus lessoni
Camaroptera chloronata granti
Figure 3. HSI maps produced by PREDICT for nine endemic Bioko birds. Open circles denote observation/mist-netting points used in the study and closed circles are those points in which the bird species was detected.
879 to be utilised even under field conditions due to the relative processing simplicity. PREDICT can be used either as a stand-alone program or as a module that can be incorporated into most GIS environments.
Figure 3. Continued.
0.79 ± 0.22 0.83
Mean ± SD Median
Ploceidae Estrildidae Sturnidae
Zosteropidae
0.41 0.83 0.48 0.95 1.00 1.00 0.62 1.00 0.84
Pogoniulus scolopaceus stellatus Camaroptera chloronata granti Batis poensis Tersiphone rufiventer tricolor Speirops brunneus Zosterops senegalensis poensis Ploceus albinucha maxwelli Estrilda nonnula elizae Lamprotornis splendidus lessoni
Capitonidae Sylviidae Musicapidae
Altitude
Taxon
Family
0.90
0.88 ± 0.07
0.91 0.85 0.84 0.87 0.90 0.90 0.95 0.98 0.71
Aspect
Probability values
0.65
0.57 ± 0.25
0.44 0.27 0.24 0.65 0.82 0.75 0.83 0.87 0.26
Slope
0.91
0.83 ± 0.27
0.73 0.10 0.86 1.00 1.00 0.91 0.91 0.95 1.00
Rainfall
0.99
0.90 ± 0.24
0.99 0.23 0.99 0.93 1.00 1.00 0.98 1.00 0.98
Vegetation
Table 1. Probability values for the degree of association between habitat factor and endemic bird distribution in Bioko.
880
881
Figure 4. Test distribution maps using PREDICT showing the actual, predicted distributions and observation points used in the testing. Percent detection rates for the different distributions were (a) 91%; (b) 97%; (c) 97%; (d) 98% and (e) 98%.
As with most GIS projects, the data used in this study were derived from a variety of sources and encompassed both differing scales and levels of accuracy. GIS enables these disparate sets of data to be brought together into a single spatial framework for analysis. Indeed, it is the ability to combine data in this manner that makes GIS so powerful. However, there are serious implications that need to be considered when combining data of any kind. All, or at the very least, most data will incorporate some degree of error. This inherent error is usually insignificant both in the context for which the data was originally produced and when viewed in isolation. Unfortunately, GIS projects are rarely able to use data produced specifically for the task. Furthermore, the actual process of combining data multiplies any errors inherent in each data set to produce a cumulative error, which may or may not be significant to the result. Generally, it is advisable to attempt to limit the potential extent of cumulative error by utilising the most accurate
882
Figure 4. Continued.
source data available. As an example, a line 0.5 mm wide drawn on 1:250,000 scale map covers an equivalent ground area of 125 m. The same line on a 1:50,000 scale map covers a ground area of only 25 m. While this example illustrates only one cartographic issue, that of line thickness, it is clear that, on the whole, more accurate maps incorporate less inherent error. This is for the reason that as scales become larger, more and more detail can be depicted until, eventually, true scale is reached, i.e. 1:1. This principle holds for all types of data, not only that in map form. Geographical analysis, and GIS in general, is an aid to decision-making. This point is often forgotten in the quest for more accurate information. It is true that in most cases better quality data will result in the limitation of cumulative error and, as a consequence, lead to more reliable results. However, increased data quality usually comes at a price. This may be in terms of time required to gather and/or process the information, the space required to store the information, the cost of the information, or more likely, a combination of the three. Therefore, once minimal data requirements have been met, a balance needs to be struck between the desire for increased accuracy and the financial resources available. Minimal data requirements themselves can be
883
Figure 4. Continued.
defined as those necessary to ensure results that are suitable for the purpose of the project. Final confirmation of the accuracy of predictions about endemic bird diversity patterns on Bioko will come to light with future field observations. Nevertheless, this method allows for areas of conservation priority to be identified and facilitates a better understanding of the relationship between species and the environment in Bioko. A major limitation common to this type of project is the precision of the environmental variable data available. This study used cartographic information derived from sources with differing levels of validity, comprehensiveness, scales and accuracy. This limiting factor affects output reliability and must therefore be taken into account when interpreting results. Nonetheless, some sacrifice of ecological precision is acceptable for the sake of the generality and usefulness of statistical predictions (Miller et al. 1989). There is one important difference between the areas assessed as being suitable and those assessed as being unsuitable. While there are no examples of a particular species being found in areas designated by the program as unsuitable, there are certain areas
884 that are considered suitable even though the species was not found there. Reasons for this could be that the species actually does exist in those areas but was not detected due to limitations in the capture or observation techniques employed, or the species is not located for some reason other than the five habitat factors considered (e.g. the existence of a predator or the absence of some other habitat feature important for its survival). The absence of a species in an area identified as being suitable can also be explained as the normal operation of metapopulation dynamics. Indeed, metapopulation models predict that a fraction of suitable habitats will be unoccupied, the fraction being a function of size and isolation of habitat patches, dispersal abilities of the focal organism and past history. Thus, factors specific to individual species could not be incorporated into a data-driven, multi-purpose prediction module such as PREDICT. This must be taken into account when interpreting the maps. The development of methods that help visualise patterns of distribution of species in little known regions, such as in our example of Bioko birds, will no doubt expedite the establishment of more realistic conservation strategies. With this in mind, PREDICT can be an effective tool to generate HSI maps of target species as well as assess future scenarios by integrating datasets of proposed developments and forecast future conditions. PREDICT can also be of use in the determination of suitable areas for re-introduction of species, perhaps even to locations where it has never previously existed. Additionally, the method can provide a useful approach for detecting suitable sites for similar species in other totally unexplored regions (e.g. using weightings for genera or species from Bioko to predict the same taxa in Mount Cameroon or the neighbouring Gulf of Guinea Islands).
Acknowledgements We wish to thank Mr David Chapman for providing us with advice and the use of computer hardware and software facilities at University College, London. Prof. David Unwin and Bob Burn provided important guidance on statistics and Dr Philip Bacon, Dr Robert A. Cheke, Dr Peter J. Jones made useful comments on the manuscript. The Cooperación Española Programme in Equatorial Guinea sponsored fieldwork in Bioko. Drs Javier Castroviejo Bolivar and Francisco J. Purroy Iraizoz provided invaluable logistical and technical assistance during the field study programme.
References Akçakaya HR (1996) RAMAS® Metapop: Viability Analysis for Stage-Structured Metapopulations (Version 2.0). Applied Biomathematics, Setauket, New York Bailey RG and Hogg HC (1986) A world ecoregions map for resource reporting. Environmental Conservation 13: 195–203 Bonham-Carter GF (1994) Geographic Information Systems for Geoscientists. Pergamon Press, Oxford
885 Butterfield BR, Csuti B and Scott JM (1994) Modelling vertebrate distributions for gap analysis. In: Miller RI (ed) Mapping the Diversity of Nature, pp 53–68. Chapman & Hall, London Fa JE (1992) Conservación de los Ecosistemas Forestales en Guinea Ecuatorial. IUCN, Gland and Cambridge Fa JE and Morales LM (1993) Patterns of mammalian diversity in Mexico. In: Ramamoorthy TP, Bye R, Lot A and Fa JE (eds) The Biological Diversity of Mexico, pp 319–364. Oxford University Press, New York and Oxford FED/DHV (1990) Proyecto de Rehabilitación de Cacao. Isla De Bioco. Uso Actual y Potencial de las Tierras. República de Guinea Ecuatorial. Ministerio de Agricultura, Ganadería, Pesca y Forestal. Unpublished document, Malabo Fish and Wildlife Service (1981) Standards for the Development of Suitability Index Models. Ecological Services Manual 103. US Department of the Interior, Fish and Wildlife Service, Division of Ecological Services. Government Printing Office, Washington, DC IGN (1979) República de Guinea Ecuatorial: Provincia de Bioco. Mapa Topográfico Escala 1: 100.000. Instituto Geográfico Nacional de España, Madrid Juste J (1992) Zonación Ecológica y Evaluación del Impacto Ambiental de los Usos Actuales de la Ísla de Bioco. Proyecto UTF/EQG/002/EQG, FAO, Malabo Juste J and Fa JE (1994) Biodiversity conservation in the Gulf of Guinea islands: taking stock and preparing action. Biodiversity and Conservation 3: 759–771 Lancia R, Adams DA and Lunk EM (1986) Temporal and spatial aspects of species-habitat models. In: Verner J, Morrison ML and John Ralph C (eds) Wildlife 2000: Modeling Habitat Relationships of Terrestrial Vertebrates, pp 65–70. The University of Wisconsin Press, Wisconsin Miller RI (1994) Setting the scene. In: Miller RI (ed) Mapping the Diversity of Nature, pp 3–18. Chapman & Hall, London Miller RI, Stuart SN and Howell KM (1989) A methodology for analyzing rare species distribution patterns utilizing GIS technology: the rare birds of Tanzania. Landscape Ecology 2: 173–189 Morrison ML, Marcot BG and Mannan RW (1992) Wildlife-habitat relationships. Wisconsin University Press, Wisconsin O’Brien LG (1989) The Statistical Analysis of Contingency Table Designs. CATMOG 51, Environmental Publications, University of East Anglia Ocaña Garcia M (1961) Factores que influencian la distribución de la vegetación en Fernando Poo. Anales del Instituto de Estudios Africanos 55: 67–85 Osborne PE and Tigar BJ (1992) Interpreting bird atlas data using logistic models: An example from Lesotho, Southern Africa. Journal of Applied Ecology 29: 55–62 Perez del Val J (1994) Contribución al Conocimiento de la Historia Natural de la Avifáuna de la Isla de Bioko (Guinea Ecuatorial). PhD thesis. Universidad de Leon, Spain Scott JM, Davis F, Csuti B, et al. (1993) Gap analysis: a geographic approach to protection of biological diversity. Wildlife Monographs 123: 1–41 Unwin DJ (1995) Map overlay in GIS as a statistical problem. In: Fischer M, Schoten HJ and Unwin DJ (eds) Analytical Perspectives on GIS in the Environmental and Socio-Economic Sciences, pp 65–70. Taylor and Francis, London Walker PA (1990) Modelling wildlife distributions using a geographic information system: kangaroos in relation to climate. Journal of Biogeography 17: 279–289
