1Many species are adversely affected by human activities at large spatial scales and their conservation requires detailed information on distributions. Intensive ground surveys cannot keep pace with the rate of land-use change over large areas and new methods are needed for regional-scale mapping.
2We present predictive models for great bustards in central Spain based on readily available advanced very high resolution radiometer (AVHRR) satellite imagery combined with mapped features in the form of geographic information system (GIS) data layers. As AVHRR imagery is coarse-grained, we used a 12-month time series to improve the definition of habitat types. The GIS data comprised measures of proximity to features likely to cause disturbance and a digital terrain model to allow for preference for certain topographies.
3We used logistic regression to model the above data, including an autologistic term to account for spatial autocorrelation. The results from models were combined using Bayesian integration, and model performance was assessed using receiver operating characteristics plots.
4Sites occupied by bustards had significantly lower densities of roads, buildings, railways and rivers than randomly selected survey points. Bustards also occurred within a narrower range of elevations and at locations with significantly less variable terrain.
5Logistic regression analysis showed that roads, buildings, rivers and terrain all contributed significantly to the difference between occupied and random sites. The Bayesian integrated probability model showed an excellent agreement with the original census data and predicted suitable areas not presently occupied.
6The great bustard's distribution is highly fragmented and vacant habitat patches may occur for a variety of reasons, including the species’ very strong fidelity to traditional sites through conspecific attraction. This may limit recolonization of previously occupied sites.
7We conclude that AVHRR satellite imagery and GIS data sets have potential to map distributions at large spatial scales and could be applied to other species. While models based on imagery alone can provide accurate predictions of bustard habitats at some spatial scales, terrain and human influence are also significant predictors and are needed for finer scale modelling.
Many wild species are adversely affected by human-induced changes in land use that operate over very large spatial scales. For example, in Europe agricultural policy change and its consequent effects on farming practice have profoundly influenced many bird species (O’Connor & Shrubb 1986; Pain & Pienkowski 1997). Among these is the great bustard Otis tarda L., a globally threatened species that has suffered dramatic declines (Collar & Andrew 1988; Heredia, Rose & Painter 1996). Although the reasons for such declines are not completely understood, it seems that agriculture intensification and habitat fragmentation due to human activities have played a decisive role. The species’ stronghold is now the agricultural landscape of Spain, with an estimated population of 20 000 birds (Alonso & Alonso 1996), more than half the world total (Hidalgo de Trucios 1990; Del Hoyo, Elliot & Sargatal 1996). The population was probably declining until hunting was outlawed in Spain in 1980 and is now thought to be stable at best. Its conservation is still threatened by habitat fragmentation over most of the Iberian peninsula. Recent dispersal studies using individual marking and radiotracking techniques have shown that although the species is capable of performing considerable seasonal migration, individuals display a marked site fidelity to their breeding areas (Alonso, Morales & Alonso 2000; Morales et al. 2000). Once traditional sites are lost, this behaviour may restrict the potential to establish new populations elsewhere.
Monitoring national or regional changes in great bustard distributions and numbers through field surveys cannot realistically keep pace with the rate of agricultural and infrastructure development. This is equally true for large numbers of other species that require assessments of the reasons for population decline. Such knowledge is essential for compiling conservation management action plans under international conventions and legislation such as the Biodiversity Convention and the Birds Directive. There is thus an urgent need to develop ways for mapping threatened species at large spatial scales with reduced field effort (Gaston & Blackburn 1995; Williams et al. 1997). In this paper we present the results from a pilot study that attempted to model the breeding distribution of great bustards in central Spain from remotely sensed data and digitally mapped data layers. Large-scale studies continue to pose major challenges in applied ecology and model development may provide ecological insight at scales where manipulation is not possible (Ormerod, Pienkowski & Watkinson 1999; Caldow & Racey 2000).
Both bustard sexes are highly aggregated in early spring, when the surveys were conducted (males usually in a single flock and females in just a few flocks). Until recently great bustards in Iberia were considered sedentary in the vicinity of breeding leks (areas for male sexual exhibition and copulation). However, work on radio-marked birds has demonstrated that both sexes behave as partial migrants between the lek site and post-breeding or wintering areas (Alonso, Morales & Alonso 2000; Morales et al. 2000) and generally show strong interannual fidelity to lek sites in spring. Females nest close to the lek where they copulate, and take over all brood caring duties. Thus surveys conducted in spring are likely to reveal consistent breeding distributions, but these may differ from wintering sites which are not addressed here.
Data availability is a constraint in building large-scale models of species’ distributions, and two basic approaches seem to be emerging to make best use of available resources. Interpolation methods, ranging from simple linear interpolation (Farina 1997) to kriging (Palma, Beja & Rodrigues 1999), estimate species’ occurrences between sample points based on their spatial arrangement. This is likely to be most successful where habitat discontinuities are few, but we know of no published studies that assess model performance. The alternative approach, which may generally be called correlative mapping, relates species’ occurrences at points to a suite of predictor variables that are available across the whole study area (Osborne & Tigar 1992; Buckland & Elston 1993; Augustin, Mugglestone & Buckland 1996). Derived equations are then used to predict occurrences across the species’ range. This is a data-hungry approach because environmental features are needed for every grid square or pixel covering the species’ distribution. However, it is likely to detect more subtle changes in distributions than interpolation methods, providing the predictor variables are reasonably correlated with the habitat features chosen by the species being mapped. Fortunately, the digital data sets now available reasonably approximate some ecological requirements of the great bustard and probably other species too. Our emphasis throughout was on the use of readily available data sets that later would permit scaling-up to the national or regional scale.
Our starting premise was that vegetation type, terrain characteristics and human disturbance determine bustard distributions in Spain, factors that may apply equally to other species. Great bustards favour open, steppe-like, landscapes comprising cereal–fallow rotations, a habitat that is particularly under threat of intensification through irrigation under European Union agricultural policy. Numerous studies (Lyon 1983; Avery & Haines-Young 1990; Austin et al. 1996; Lavers, Haines-Young & Avery 1996) have related bird distributions to habitats using remotely sensed imagery such as LANDSAT thematic mapper (TM) and multispectral scanner (MSS), but the high cost precludes their use over extensive areas. Meteorological satellite data from the advanced very high resolution radiometer (AVHRR) operated by the National Oceanic and Atmospheric Administration are much cheaper and more readily available, but suffer from coarse spatial resolution (c. 1 km) which may limit discrimination of land-use types. However, the high temporal resolution of these data can offer an alternative route. Several workers (Kremer & Running 1993; Reed et al. 1994; Paruelo & Lauenroth 1995) have utilized time series of AVHRR normalized difference vegetation indices (NDVI) to map and monitor a range of habitats. NDVI is calculated from the near infra-red (NIR) and red (R) spectral bands as NDVI = (NIR − R)/(NIR + R), exploiting the fact that vigorous vegetation reflects strongly in the NIR and absorbs radiation in the red band (Mather 1999). Rogers & Williams (1994) and Rogers et al. (1997) have taken this approach one step further by using NDVI to discriminate wildlife habitats, while others (Walker et al. 1992; Fjeldsa et al. 1997) have utilized the same data for identifying and understanding biogeographic patterns. We used a short time series of AVHRR data to predict the vegetation component of great bustard distribution in central Spain.
Few studies have actually demonstrated a preference for flat to slightly undulating terrain (although see Alonso & Alonso 1990; Onrubia et al. 1998), but it is a commonly stated habitat requirement of great bustards (Johnsgard 1991). The birds probably prefer sites with good horizontal visibility both to watch for predators and because the breeding system of dispersed leks involves strong visual cues over long distances. In our study region the species is distributed around 11 leks. Between late winter and early spring, males concentrate at these traditional arenas where they display and fight to establish a hierarchical rank. Females also gather at these arenas to mate between late March and early April. If lekking bustards select certain terrain characteristics, these should be discernible in currently available digital terrain models (DTM) with resolutions under 100 m.
Evidence for an effect of human disturbance on bustards is similarly anecdotal, although it is generally assumed that the presence of human infrastructures affects the distribution negatively. However, Lane, Alonso & Martin (2001) reported the absence of flocks within a band of about 1 km around villages and busy roads. For other birds, strong effects of roads on breeding density and performance have been noted (Reijnen & Foppen 1994; Reijnen et al. 1995). Analysis of the effects of roads and buildings on distributions should be straightforward with any accurate vector data source.
The pilot study area measured c. 126 × 132 km, centred on Madrid province, Spain, with the lower right co-ordinate at 2°44′ W, 39°49′ N (Fig. 1). It was chosen because accurate great bustard census data, geographic information system (GIS) data coverages and satellite imagery were readily available. The area comprised 55·9% agricultural land, 29·2% natural scrub or wooded cover, 8·2% forestry, 5·6% built environment and the remaining 1·1% bare ground or open water (European Union Corine Land Cover Project). The majority of analyses were confined to Madrid province itself because this was where bustard censuses were conducted but, where possible, models were extrapolated to the full study area (see later).
Digitized infrastructure maps were available from Autonomous Community of Madrid Cartographic Service at 1 : 100 000 scale. The data were separated into four layers (roads, buildings, railways and river systems) using ArcView software (ESRI 1996) and then rasterized to 80-m pixels in Idrisi (Eastman 1995). This resolution was chosen because a digital terrain model (DTM) was also available at this scale for the province. We created new variables from each of the infrastructure layers by replacing the central pixel of a 13 × 13-cell moving window with the proportion of pixels recording the feature of interest (Table 1). This is equivalent to calculating the percentage land cover of the feature at 80-m resolution within a c. 1-km2 quadrat.
Table 1. Predictor variables used to compare pixels occupied by bustards with random locations
Proportion of 80-m pixels in a 13 × 13-array containing roads. Equivalent to the density of roads at 1·1-km resolution
Proportion of 80-m pixels in a 13 × 13-array containing buildings or large built structures such as airfields
Proportion of 80-m pixels in a 13 × 13-array containing railway tracks
Proportion of 80-m pixels in a 13 × 13-array containing rivers
The altitude in m recorded on the 80-m digital terrain model
Terrain variability 25
Coefficient of variation in altitude in a 5 × 5-pixel array of 80-m pixels. Measures variation in altitude at 0·16-km2 resolution
Terrain variability 81
Coefficient of variation in altitude in a 9 × 9-pixel array of 80-m pixels. Measures variation in altitude at 0·52-km2 resolution
Terrain variability 169
Coefficient of variation in altitude in a 13 × 13-pixel array of 80-m pixels. Measures variation in altitude at 1·1-km2 resolution
% slope. The maximum of either the north–south or east–west slopes across a 3 × 3-pixel array
The value of the normalized difference vegetation index for each month based on a maximum value composite of AVHRR imagery at 1·1-km2 resolution. Scaled 0–255
The DTM was used to derive the altitude of each pixel and to calculate terrain variability as bustards appear to choose open, gently undulating, landscapes. Using moving windows of 5 × 5, 9 × 9 and 13 × 13 pixels we calculated the coefficient of variation (CV) in altitude and placed this value in the central pixel. The three window sizes were selected to examine the effect of scale on terrain variability.
A range of cloud-free AVHRR images at 1·1-km resolution was obtained for each month during 1996 to calculate the temporal NDVI signature for each pixel. Several factors can affect the reliability of NDVI values extracted from these data: changing illumination and viewing conditions within a single image and between images on different days; the presence of cloud cover; variations in atmospheric constituents such as water vapour and aerosols (Marçal & Wright 1997). The use of maximum value composites (MVC) is a means of partial correction of AVHRR data for the effects of these different factors (Holben 1986). Given a large number of images throughout a year, the assumption behind this approach is that the maximum NDVI values of the image set will correspond to ideal conditions, i.e. low solar zenith and viewing angle, low water vapour and aerosol concentrations and cloud-free conditions (Marçal & Wright 1997). However, for time series generated over large spatial areas, the MVC approach can have some limitations, as variations resulting from off-nadir viewing geometry cannot always be accommodated or corrected (Stoms, Bueno & Davis 1997). For this study, which concentrated on a relatively small study area, viewing geometries for MVC data were assumed to be near-nadir. NDVI values were therefore calculated for each image and used to form a MVC time series from the best cloud-free images from each month.
Bird census data
A complete great bustard census was conducted in March 1997 for the whole of Madrid province. The distribution of the species in the study area was known from previous censuses (Alonso & Alonso 1990, 1996). During 1 week three teams each of two experienced observers counted bustard flocks at the known breeding areas and also searched all other potential sites. In practice, birds sighted just beyond the province boundary were also recorded. One-hundred and three flocks (960 birds) were first marked on field maps, then digitized and the point coverage rasterized to 80-m and 1·1-km resolutions, recording the presence of bustards in the pixel. This resulted in 71 pixels with one or more flocks at 1·1-km resolution, and 92 pixels at 80-m resolution. For comparison, we generated equivalent random point coverages, stratifying them geographically both to sample the whole province and to reduce spatial autocorrelation (i.e. to reduce the probability of using adjacent pixels). The number of random points selected is important because prevalence (i.e. the ratio of positive to negative pixels) affects the outcome of model performance testing in logistic regression as used here (Fielding & Bell 1997; Manel et al. 1999).
Analyses were based on the comparison of landscape features at the random points and at the locations used by bustards. In grid-cell mapped data the appropriate test to use depends on the spatial autocorrelation in the sampling points because this in effect reduces the degrees of freedom and thus increases the chance of type I errors (Cliff & Ord 1981; Legendre & Legendre 1998). We assessed spatial autocorrelation using Moran's I and then for univariate analyses followed the advice of Cliff & Ord (1981) for modifying t-tests based on the results.
Multivariate analyses were carried out using forward stepwise logistic regression (SPSS 1997) to contrast pixels used by bustards with the random set. Analyses were carried out separately on the GIS data at 80-m resolution and the AVHRR imagery at 1·1-km resolution (Table 1). For the latter, models included the 12-monthly NDVI values and two NDVI contrasts (April minus July, and April minus January) derived empirically by inspection of mean NDVI temporal signatures (Fig. 2). We also included quadratic terms for the predictor variables to allow for the possibility of optimum vegetation conditions being selected by the birds (quadratic terms within logistic regression model Gaussian responses).
As in univariate analysis, spatial autocorrelation affects significance tests on logistic regression coefficients and, as no satisfactory method exists at present to correct for this, caution is needed in their interpretation. Conventional statistical modelling on spatial data ignores spatial autocorrelation in the residuals due to the ecological likelihood that neighbouring pixels will have dependent probabilities of use. To overcome this, we adopted the approach of Augustin, Mugglestone & Buckland (1996) by incorporating an autologistic term in the models based on the modified Gibbs sampler. A probability surface was first generated by logistic regression in the usual way. Then a moving window of 9 × 9 pixels was used to calculate the mean of the probabilities assigned to the 80 neighbouring cells, weighted by Euclidean distance. This autologistic term was entered into the regression and the model rerun. The procedure is iterated to stability to produce the final probability surface (for further explanation see Augustin, Mugglestone & Buckland 1996). In image-processing terms, the autologistic term acts as a smoothing filter, removing isolated pixels and consolidating habitat patches defined as suitable.
Results of logistic regression models are often judged as successful if predicted probabilities > 0·5 correspond with observed occurrences and values < 0·5 with absences. However, this dichotomy is arbitrary and lacks any ecological basis; patches rated with a 0·6 probability of occurrence may in fact be unsuitable. The more powerful approach used here is to assess model success across the full range of dichotomies using receiver operating characteristics (ROC) plots. ROC plots are widely used in clinical chemistry (Beck & Shultz 1986; Zweig & Campbell 1993) but rarely by ecologists (Fielding & Bell 1997). A ROC plot depicts on the y-axis sensitivity, i.e. a/(a + c) in a 2 × 2 confusion matrix of the model prediction against the observations. This is plotted against 1 − specificity, i.e. 1 − (d/(b + d)) from the same confusion matrix. The chance performance of a model lies on the positive diagonal of a ROC plot, whereas models that out-perform chance follow a curve lying in the upper left half. The area under the ROC curve (AUC) is a convenient measure of overall fit and varies from 0·5 (for a chance performance) to 1·0 for a perfect fit. We generated ROC plots using SPSS software and calculated the AUC and its standard error using a non-parametric approach. The results are reported here as the AUC ± its standard error together with the significance of a test that the area = 0·5, i.e. that the model results do not differ from chance.
The probability surfaces derived from the two independent logistic regression models were combined using Bayesian inference after first resampling the 1·1-km surface to 80-m resolution. The technique permits prior probabilities (for example derived from one model) to be revised on the basis of new probabilities calculated from a second model. The appropriate formula (from Pereira & Itami 1991) is:
where PNDVI is the probability derived from the NDVI model, and PGIS is the probability derived from the model based on GIS data layers. Bayesian approaches to decision-making have previously been used as here by Pereira & Itami (1991), and in other ways in wildlife distribution modelling by Aspinall & Veitch (1993) and Tucker et al. (1997).
At 80-m resolution, neither the bustard locations nor the random points exhibited significant spatial autocorrelation (Moran's I < −0·0001 in both cases). Following Cliff & Ord (1981) we therefore compared site characteristics using standard t-tests adjusted for unequal variance.
Sites occupied by bustards had significantly lower densities of roads, buildings, railways and rivers than random points (Table 2). Of these, the effect of buildings was particularly strong; bustards occurred at sites with a mean of only 0·8% land cover by buildings, whereas random sites averaged 10·5% (P < 0·001). Table 2 gives the ranges for each variable within which bustards occurred and these are combined as a threshold mask in Fig. 3a. This clearly shows the elimination of the radiating network of roads and buildings from Madrid City.
Table 2. Comparison of features around 92 sites occupied by great bustards and 93 random points (except for terrain variability and slope based on 83 and 87 sites, respectively, to eliminate edge effects). Values are means ± standard deviations and all t-tests are adjusted for significant unequal variance
Range used by bustards
0·022 ± 0·045
0·043 ± 0·080
0·008 ± 0·026
0·105 ± 0·216
0·006 ± 0·022
0·018 ± 0·052
0·028 ± 0·044
0·053 ± 0·066
673·9 ± 56·9
799·7 ± 285·4
Terrain variability 25
0·540 ± 0·413
0·949 ± 0·814
Terrain variability 81
0·871 ± 0·555
1·518 ± 1·027
Terrain variability 169
1·153 ± 0·649
1·996 ± 1·270
3·08 ± 3·05
6·60 ± 7·92
Bustards occurred within a narrow range of elevations from 566 to 780 m a.s.l., whereas random points covered the range 520–2194 m a.s.l. The terrain surrounding bustard sites was also significantly less variable than that around random sites at all three scales examined (Table 2). There was no obvious trend for a difference between the scales, although the significance of the difference between bustard and random sites increased within window size. Using 13 × 13 cells (about 1 km2), bustards occurred at sites with up to 2·8% coefficient of variation in altitude (mean 1·2%), whereas random sites had up to 5·9% variability (mean 2·0%). Combining the results for altitude and terrain variability with the mask in Fig. 3a yields Fig. 3b. This indicates that only 2133 km2 of the 6540 km2 (32·6%) studied meets the characteristics of infrastructure, elevation and river networks at sites used by bustards. Furthermore, this area is fragmented into blocks by the radiating networks from Madrid City.
Predictive modelling using logistic regression
Building threshold masks (e.g. Fig. 3) is a convenient way to define areas meeting criteria but suffers from failing to take into account interactions between variables. Using the GIS variables in Table 2 (selecting terrain variability 169 for variation in elevation), we built predictive models for bustard presence using forward stepwise logistic regression. Only railways and slope were not included in the model (at P < 0·05) and all other variables were significant at P < 0·01 (Table 3). The greatest contributions came from terrain variability (P < 0·0001) and housing density (P < 0·0002). Overall the ROC plot for the model (Fig. 4a) had an AUC of 0·898 ± 0·023 and was highly significant (P < 0·001). A simplified probability surface for bustard occurrence based on the significant GIS variables is shown in Fig. 5. There are obvious similarities with Fig. 3b but the probability plot shows far more texture and greater weighting in the east-central and extreme south-west parts of the province.
Table 3. Summary results of the logistic regression analyses. The significance of the coefficients was assessed using the Wald statistic
GIS data layers (80-m resolution)
Terrain variability 169
NDVI data (1·1-km resolution)
NDVI (April) – NDVI (July)
We built a separate model based on the satellite data at 1·1-km resolution. The stepwise inclusion of the explanatory variables resulted in the selection of three NDVI variables for January, November and the contrast April minus July (Table 3). None of the quadratic terms was significant. The model ROC plot (Fig. 4b) had an AUC of 0·93 ± 0·022 and was highly significant (P < 0·001). The resultant probability surface is shown in Fig. 6. Although this model of habitat suitability was derived by analysing only the presence or absence of bustards, the probabilities also relate to the numbers of birds present (Fig. 7). Most points lie beneath the diagonal from bottom left to top right, indicating that large numbers of birds only occur in pixels with high predicted probabilities.
We took account of autocorrelation in the model based on the satellite data by including an autologistic term; the regression coefficients stabilized within five iterations to produce Fig. 8. Comparing this with Fig. 7 reveals the down weighting of isolated pixels previously defined as suitable, and the consolidation of the larger suitable habitat blocks. The ROC plot for the autologistic model (Fig. 4c) differed very little from the standard NDVI model (AUC = 0·93 ± 0·022, P < 0·001). The 1·1-km2 resolution probability surface in Fig. 8 was resampled to 80-m resolution in order to combine it with the thresholds imposed by infrastructure and natural features (i.e. Fig. 3). Taking an arbitrary lower threshold of 50% probability of occurrence from the habitat suitability map (Fig. 8) alone yielded an area of 972 km2 or 14·9% of the province. This provides an estimate of the area of the province with vegetation index signatures comparable to those used by bustards. When combined with Fig. 3b, however, this falls to 570 km2 or 8·7% of the province (Fig. 9). The difference between these two values (400 km2) is the amount of apparently suitable habitat that bustards are unlikely to use due to proximity to unfavourable landscape elements. Figure 9 (area 570 km2) is the area of Madrid province that cannot be distinguished from sites occupied by bustards using GIS and remote sensing, based on simple threshold mapping.
An alternative approach is to combine the two probability surfaces based on feature data (Fig. 5) and NDVI (Fig. 8) using Bayesian integration (see the Methods). As Fig. 5 represents largely fixed features of the landscape today, we regarded these as prior probabilities that may be refined by consideration of Fig. 8 which represents land cover based on green biomass. The Bayesian integrated probability model (Fig. 10) showed an excellent agreement with the original census data (simplified in Fig. 11 for ease of display) and had an ROC AUC value of 0·969 ± 0·013, P < 0·001 (Fig. 4d).
We started with a premise gleaned from the literature that great bustard distributions may be related to vegetation, topography and human influence, and sought digital data sets that could characterize these features. At the landscape scale, our analyses successfully predicted the occurrence of bustards around Madrid province and all these features were significant predictors.
The negative effect of human disturbance on bustard occurrence is not surprising but is interesting because at first sight it appears that the birds are well integrated into the urban setting in Madrid province. There is growing evidence that roads impact on large mammals (Mech 1989; Mladdenoff et al. 1995; Mace et al. 1996; Palma, Beja & Rodrigues 1999) and birds (Reijnen & Foppen 1994; Reijnen et al. 1995) and this is important in the context of environmental impact analysis (Hill et al. 1997). The negative effect of buildings was even stronger and independent of the impact of roads in our multiple logistic regression analyses. A similar analysis on grey geese (Anser anser and A. brachyrhynchus) in Scotland found the same result (C. Urquhart & P.E. Osborne, unpublished data) and it may be important that future studies consider the possible effects of roads and housing independently.
We do not believe that the negative impact of rivers and their tributaries demonstrates an actual avoidance of these features but rather that irrigated crops are frequent along watercourses and these are avoided. Bustards occupied only a part of the altitude range available and, more significantly, occurred at sites where the surrounding terrain was less variable than at random sites. This accords with their assumed preference for relatively open terrain. Our analysis with three window sizes (400 m to 1 km) did not detect the point where terrain variability at bustard and random sites became the same, although logically with a large enough window this would be true. This suggests an avenue for further research; indeed, finding the scales at which impacts of impaired visibility and disturbance are greatest would provide valuable data for conservation.
As Fig. 2 illustrates, there was a marked difference between the mean NDVI signatures for sites used by bustards and those that were not. Although AVHRR data are relatively coarse-grained, the use of temporal signatures provided enough resolution to discriminate occupied and unoccupied sites and indicates the potential of using NDVI time series to predict avian habitats. We believe that our model largely detected the sharp change in NDVI brought about by rapid crop growth in cereal fields in spring followed by biomass loss through harvesting in July. However, based on local knowledge, it appears that the model is not simply predicting all cereal areas, because some large areas were excluded. How the subset of more suitable fields is recognized is less clear. Maurer (1994) has previously shown a relationship between avian abundance and NDVI, but the correlation of abundance with probabilities of occurrence generated by temporal NDVI curves is new. The relationship here indicates that sites defined as suitable have the potential to hold most birds, although they do not necessarily do so. Conversely, sites defined as unsuitable appear capable of holding only few birds.
Despite the coarse-grained nature of AVHRR data, models based on NDVI alone do provide significant predictions of bustard occurrence. In some situations, remotely sensed data may be all that are available and our results indicate that they are a useful first step in model building. However, we conclude that while models based on vegetation alone can provide accurate predictions of bustard habitats at some spatial scales, terrain and human influence are also significant predictors and are needed for finer scale modelling. For example, the breeding site in the centre of the south-east part of the province (Fig. 11) was not predicted by the NDVI probability surface (Fig. 8) but appeared after integration with the landscape GIS variables. We thus concur with Manel et al. (1999) that prediction success in distribution models may be enhanced when local data are available, despite the apparent success of coarse-grained models.
The use of an autologistic term based on a modified Gibbs sampler (Augustin, Mugglestone & Buckland 1996) in the logistic regression models proved very effective in sharpening the definition between occupied and unoccupied patches. Although similar visual effects can be achieved more simply in image processing through Gaussian or median filters, they lack an ecological basis, whereas the Gibbs sampler uses information on the proximity of occupied pixels. We chose to consider neighbours within 4·4 km of the target pixel (i.e. a 9 × 9 window at 1·1-km resolution) because this seemed an appropriate scale for breeding great bustards, comparable with the average area of influence of a lek. Augustin, Mugglestone & Buckland (1996) tried a range of window sizes to 4 km distance for red deer Cervus elaphus but found that 3 km gave optimum results in terms of the balance between model accuracy and computation time. Further work is needed on the selection of optimum neighbourhood distances for a range of species when trying to account for spatial autocorrelation in distribution models.
Generally, the prediction of occupied sites was more successful than the prediction of absences. For example, the Bayesian integrated model (Fig. 10) correctly predicted 93·0% of occupied sites and 78·9% of unoccupied sites, based on a 50% probability dichotomy. This is the opposite result to that reported by Manel et al. (1999) for six bird species and reinforces their view that more work is needed on model performance indicators and their determinants. Differences in performance at predicting presence and absence may be due to a number of factors (Fielding & Bell 1997) but we offer the following ecological explanations. First, the bustard census data were a single-day snapshot of occurrences, taken when flocking was at its maximum at lek sites. Under these conditions, adjacent areas could be mistakenly regarded as unsuitable whereas with greater flock dispersion they would be occupied. Secondly, the Spanish great bustard population is fragmented into numerous habitat patches and, due to high site fidelity (Alonso et al. 1995), movement between patches for breeding may be scarce. Tilman, Lehman & Kareiva (1997) predict that these conditions will lead to the non-occupancy of some suitable sites even at population equilibrium. Thus we might expect some unoccupied sites to be ecologically inseparable from occupied sites. Thirdly, recent field studies by Lane, Alonso & Martin (2001) could not discriminate occupied from unoccupied but apparently suitable areas for bustards in central Spain, and it is unreasonable to expect large-scale models to outperform intensive field investigations. Their study compared 13 areas occupied by great bustards with a matched set of 12 unoccupied areas nearby. These sites could not be distinguished using discriminant analysis on a set of variables defining relevant habitat characteristics such as crop type, substrate heterogeneity, field size, presence of roads, villages and powerlines. Both their study and ours indicate that not all potentially suitable areas are occupied and that great bustards show fidelity to sites regardless of the availability of suitable habitat elsewhere. We believe that this may be due to a combination of a series of local extinction processes in recent decades due to human-induced habitat deterioration and hunting, and the very low re-colonization capability of the species which arises from its complicated lek breeding system. Settlement patterns are probably determined by the presence of conspecifics rather than habitat cues (J.C. Alonso, unpublished data). This means that conservation efforts must be directed towards protecting traditional lek sites, and that once a lek is extinct the site will probably remain empty in the future.
The limited number of historically documented extinctions (e.g. the small patch predicted as suitable west of Madrid city was occupied 20 years ago; Figs 10 and 11) and circumstantial evidence (e.g. place names in the larger area with high probability of occurrence south of the city) suggest that Fig. 11 shows the likely distribution of great bustards some 20–40 years ago. Since then many of the predicted areas have become vacant due to local extinction processes. It is also important to remember that the locations in Fig. 11 were based on breeding birds and that these perform seasonal movements and expand the occupied area throughout the year. Conservation based on breeding sites alone would almost certainly be ineffective, and a landscape-scale approach incorporating seasonal changes in locations is required.
Our study was preliminary in developing methods for landscape-scale models but the intention is to examine distributions at the national scale. We envisage two major challenges in relation to scaling-up to larger study areas. First, as mentioned earlier, the selection of maximum value NDVI data may not lead to adequate standardization over large areas due to a level-of-view and solar-angle bias. In practice, this means that a given surface may generate a different NDVI–MVC score at different locations on the image. The main factor causing this is the inherent variability of surfaces at different look angles. A traditional solution to this problem is to first standardize images against a common target such as a large body of water or bare ground (Sannier et al. 1998). A better approach, however, may be to take into account the bidirectional reflectance distribution function (BRDF) of the surfaces being sensed (Cihlar, Manak & Voisin 1994). Correction for these effects can be undertaken using linear semi-empirical kernel-driven models (Roujean, Leroy & Deschamps 1992) that adjust image data for BRDF variability and extract surface information. Initial results indicate that the new models provide enormous scope for improving data quality and also provide useful additional information (Chopping 1998).
The second challenge to scaling-up is that animals may not choose habitats according to absolute needs but may adopt a comparative approach. Particularly on the edge of a species’ range, occupied sites are likely to be far from the ideal habitat with poorer breeding performance to match (Lawton et al. 1994). Also, several habitat types with different spectral signatures may be equally suitable at large geographical scales. Large-scale models must therefore include samples from across the species’ range and analysis may be better partitioned spatially prior to modelling.
We are grateful to Dr Antonio Yagüe of Infocarto, S.A., for supplying the NOAA data, and the Autonomous Community of Madrid Cartographic Service for the GIS coverages. John McArthur and Alison Martin helped prepare the data for analysis. Dr Enrique Martín, Dr Manuel Morales, Carlos Martín, Dr Simon Lane and Dr Javier A. Alonso assisted with the collection of bird data. This research was jointly funded by British Council/Ministry of Education and Culture Accion Integrada HB96-82 and HB97-69 awards, DGICYT project PB94-0068, and by the Dirección General del Medio Natural of Madrid Community.
Received 3 February 2000; revision received 29 November 2000