The effect of climate change on tropical rainforest vegetation pattern

The effect of climate change on tropical rainforest vegetation pattern

Ecological Modelling 145 (2001) 211– 224 The effect of climate change on tropical rainforest vegetation pattern Ber...

391KB Sizes 0 Downloads 41 Views

Ecological Modelling 145 (2001) 211– 224

The effect of climate change on tropical rainforest vegetation pattern Bertram Ostendorf a,b,*, David W. Hilbert b, Mike S. Hopkins b a


Adelaide Uni6ersity Waite Campus, Dept. of Soil and Water, PMB1 Glen Osmond, SA 5064, Australia Rainforest Co-operati6e Research Centre, CSIRO Tropical Forest Research Centre, P.O. Box 780, Atherton, Queensland 4883, Australia Received 25 October 2000; received in revised form 15 June 2001; accepted 29 June 2001

Abstract The effect of climatic change on tropical vegetation is of global and regional concern because of the high biodiversity and the potential feedback to the carbon, water, and nutrient cycles. One of the most critical aspects for assessing broad-scale consequences of climate change is our understanding of how vegetation may change. Models relating vegetation and environmental conditions can be developed for large regions. For a simple application of static models of vegetation–environment relationships, one would have to assume that the probability of species (or vegetation) occurrence conditional on environmental conditions is constant in time (abbreviated as the POCEC assumption). This assumption is critical and difficult. In this paper, we evaluate how the spatial arrangement of forest pattern may constrain vegetation change as predicted by a spatially static artificial neural network (ANN) model. We have relaxed the POCEC assumption by subjoining a spatially dynamic component based on the cellular automata approach. The ANN model quantifies a most suitable forest type based on the conditional probability of vegetation in the environmental space, whereas the cellular automata model imposes spatial constraints on the transition to the best-suited type. We adapt the cellular automata algorithm to successively increase spatial constraints, hence relaxing the POCEC assumption. Our study area is located in Northern Queensland and encompasses 20 000 km2. We evaluate the effect of the + 1 °C mean annual temperature and the −10% mean annual precipitation change. A comparison of predictions of vegetation change with the different models indicates that the spatial arrangement of vegetation in the ‘Wet Tropics’ region may impose relatively few constraints for the region’s potential change. Depending on the strength of spatial effects included in the models, the predicted future vegetation patterns differ from 1 to 10% of the study area. However, if in addition to spatial constraints ecological constraints also are considered (e.g. prohibiting several transitions that would appear very unlikely to experienced forest researchers), the predictions may differ by as much as 27%, showing a relatively strong dependence of predictions on assumptions about patch-level processes. Furthermore, using different models allows us to assess the uncertainty associated with predictions. The results demonstrate a relative certainty of a predicted decrease of notophyll rainforest types and an increase of medium open forests and woodlands, respectively, whereas the predictions of mesophyll vine forest and wet sclerophyll vegetation differ strongly among different models. © 2001 Elsevier Science B.V. All rights reserved. Keywords: Palaeotropics; Queensland; Species dispersal; Structural types; Classification; Vegetation modelling; Cellular automata; Artificial neural networks; Regional scale * Corresponding author. Tel.: + 61-8-830-37317; Fax: + 61-8-830-36717. E-mail address: [email protected] (B. Ostendorf). 0304-3800/01/$ - see front matter © 2001 Elsevier Science B.V. All rights reserved. PII: S 0 3 0 4 - 3 8 0 0 ( 0 1 ) 0 0 3 9 2 - 1


B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

1. Introduction The effect of climatic change on tropical vegetation is of global and regional concern because of the typically high biodiversity and the potential feedback to the carbon, water, and nutrient cycles (Bazzaz, 1998). The humid tropical region of North East Queensland is a unique area in which to study the regional effects of climate change on tropical forests because it encompasses a large climatic range of 16– 25 °C annual mean temperature and 1000–8000 mm mean annual precipitation. The area is confined to 20 000 km2 bordered by the Pacific to the east and dry woodlands to the west. The ‘Wet Tropics’ region conveys a long tradition of environmental research (for example, Webb, 1959, 1968, 1978) and allows for a well-established structural– environmental typology and detailed maps (Tracey and Webb, 1975; Tracey, 1982). Models are essential tools to assess the potential response of vegetation to climate change particularly if large spatial and temporal scales are considered. Several major lines of modelling strategies have been developed in the literature, all of which are difficult to apply in tropical rainforest regions. Gap-based or forest succession models (for example, West et al., 1981; Shugart, 1984; Bugmann, 1996; Bolliger et al., 2000) have been successfully applied in temperate regions where growth parameters could be identified for all major species. However, for tropical regions, where the current level of understanding of the ecosystems does not even allow an accurate estimate of the regional pattern of tree diversity, it is unrealistic to assume that parameterisations at the species level could be obtained, even when a simplified structure is suggested as in Chave (1999). Shugart et al. (1980) identified a parameter set for one of the vegetation types also occurring in our study region but an application for a large region seems impossible (see also Loehle and LeBlanc (1996) and Peng (2000) for a review and critique). The details needed to run these models at a spatial resolution and extent that would be needed for regional management will not be available in the near future. The same argument holds for models based on leaf and canopy-scale ecophysiological

principles (for example, Ostendorf, 1995, 1996; Ostendorf et al., 1996) or any process-based models requiring a parameterisation of flow rates between compartments (see Kickert et al. (1999) for a comprehensive review). A different, more empirical approach is based on identifying relationships between biological patterns and environmental conditions that can then be used to infer potential temporal dynamics. This methodology has the advantage that large sample sizes can be obtained through mapping or remote sensing. At the global scale, the most prominent example is the temperature/precipitation dependence of biomes quantified in the Holdridge Triangle (Holdridge, 1947; see Emanuel et al. (1985) for a global change application). The same approach has been successfully applied in regional studies (Austin, 1998; Brzeziecki et al., 1995; Huntley et al., 1995; Franklin, 1998; Kienast et al., 1998; Gottfried et al., 1998; Guisan et al., 1998; Leathwick, 1998; Ostendorf and Reynolds, 1993, 1998; reviewed in Franklin (1995) and recently in Guisan and Zimmermann (2000)). All model approaches are based on some form of empirical data. Model applications require that the fitted functional relationships in the models are general and will hold under changed environmental conditions. Models using a space for time substitution are based on rich and detailed spatial information and have a very high level of empiricism. Consequently, this assumption is most critical for this type of model. In order to use empirical spatial models for predictions, one has to assume that the probability of species (or vegetation) occurrence conditional upon environmental conditions is constant in time (abbreviated as the POCEC assumption). Note that this statistical relationship differs from assuming an equilibrium between environmental conditions and patterns. The equilibrium formulation would require that both current and future distributions of species or vegetation types are in equilibrium with their environment, which, given the magnitude of past climate changes (glaciation cycles, etc.) or disturbance (fire, severe storms, etc.) will not be valid for the majority of ecosystems. The POCEC assumption is more precise in that the ‘potential’ or

B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

equilibrium distribution of spatial pattern is beside the point. Rather, one has to carefully evaluate whether the ‘actual distribution in the environmental space’ may stay the same during both observation and prediction time periods. This means that also types influenced by disturbance can be used in models. Seral types co-exist at identical environmental conditions as later successional types, and therefore ‘equilibrium models’ would not allow their predictions. In contrast, models making the POCEC assumption would explicitly account for a probability of coexistence. The POCEC assumption is possibly the most critical and vulnerable part of predictive vegetation and habitat modelling. The validity of the POCEC assumption for future scenarios depends on biotic factors causing a time lag (dispersal, recruitment, etc.) and spatial barriers that constrain plant establishment and survival at a new location. Hilbert and van den Muyzenberg (1999), Hilbert et al. (2001), Hilbert and Ostendorf (2001) have developed and applied a model using an artificial neural network (ANN) to parameterise vegetation–environment relationships for the region. They found a high sensitivity of spatial patterns of environmental suitability for the different forest types to climate change. The question remains whether spatial constraints in the region can prohibit of modify change. In other words, may the POCEC assumption lead to errors because of spatial constraints? Cellular automata have been successfully applied to a broad range of applied and theoretical ecological questions (for example, Carey, 1996; Jeltsch and Wissel, 1994; With and King, 1998; Wiegand et al., 1998). The spatially dynamic model applied here is a linkage of the ANN model (representing an ideal suitability) with a cellular automata approach. In other words, rather than allowing a vegetation type to directly assume a new state as predicted from the ANN model, we first consider spatial constraints that might act upon transitions. We compare the results from the basic ANN model (making the POCEC assumption) with predictions of the spatially dynamic model to address the question of how the spatial arrangement of


forest types may limit the movement of forest type boundaries and how transitions may be constrained by a suite of orographic, anthropogenic, biological, and environmental factors. Our analysis is conducted using a most likely climate change scenario for doubled atmospheric CO2 in the area (+1 °C, − 10% precipitation; Walsh et al., 2000).

2. The study region The study region is located between 15–19° latitude and 145–146° longitude (Fig. 1). The area includes the largest continuous area of rainforest in Australia. The climate zones in the region range from dry in the western part to wet at the coast with highest rainfall in the coastal mountains. Long-term average annual precipitation shows a strong topographically influenced variation from below 1000 mm at the margin of the dry country, 2000–3000 mm at the coast, and up to 8000 mm on the mountain peaks. The annual average temperature reaches 26 °C in the dry country, 24 °C at the coast, and ranges from 16 to 20 °C in the mountains and tablelands, respectively. Spatial climate variables were generated by the ANUCLIM software (McMahon et al., 1995) using climate information collected by the Bureau of Meteorology and digital terrain data. The terrain grid with a pixel resolution of 1 ha was generated from 50 m (20 m in flat coastal areas) contour lines (base maps of 1:50 000). The area encompasses 2.04 million pixels, of which 16% have been mapped as cleared for agriculture (Tracey and Webb, 1975). The pixel size of 1 ha is a compromise between the resolution required to show important vegetation features such as gallery forests on the one hand and computational limitations on the other. Rainforest types were mainly distinguished by their structural features (see Webb, 1959, 1968, 1978; Tracey and Webb, 1975 for classification rationale and mapping details). Webb (1978) and Tracey (1982) present a hierarchical field key to identify the structural types. At the highest level, the most common leaf size is distinguished, fol-


B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

lowed by forest structural complexity (vines, epiphytes, plant buttresses), canopy height and the occurrence of deciduous canopy emergents or palms. Vegetation in the area has been mapped at

a scale of 1:100 000 based on aerial photographs (Tracey and Webb, 1975). The original map distinguished 23 forest classes (type 24 denotes the cleared areas) and a review of the southern part of

Fig. 1. Location of the study area. The darker shading indicates the current extent of tropical rainforest, the lighter grey shows the total extent of the mapped and analysed area.

B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224


Table 1 Vegetation classes used in the analysis, abbreviations used, and the corresponding types in the Tracey and Webb classificationa Abbreviation

Original classes




1 and 2

Mesophyll vine forests



Mesophyll vine forests with palms



Semideciduous mesophyll vine forest



Complex notophyll vine forests



Notophyll vine forests


8, 9, 10

Simple notophyll and simple microphyll forests and thickets



Deciduous microphyll vine thicket


12, 13

Vine forest with Acacia and/or Eucalyptus



Tall open forest and tall woodland



Medium open forest and woodlands


16 17–20, 22, 23

Medium and low woodlands Coastal complexes



Mountain rock pavements



Araucarian vine forests


26, 27

Notophyll semi-evergreen vine forests

Rainforest with a complex structure, high diversity of vascular epiphytes, and a canopy dominated by mesophyllous species Rainforest where the canopy is dominated by palms, occurring on poorly drained soils near the coast Mesophyll vine forest with canopy emergents often being deciduous Rainforests of cooler uplands, structurally complex, canopy dominated by notophyll species Rainforest found in drier coastal zones, simple structure, low canopy Diverse group of mountain rainforests in the coolest and wettest parts of the study area, structurally simple Rainforest with low canopy of mainly drought deciduous species Rainforest with sclerophyll canopy emergents, successional communities Sclerophyll forests with high canopies, occurring in moist environments Medium height sclerophyll forests, including poorly drained coastal locations Dry, open sclerophyll woodlands Large variety of fine grained vegetation mosaics of several rainforest and sclerophyll classes, occurring near the coast Fine grained mosaic of dry rainforest and sclerophyll classes on steep mountain slopes with thin soils Rainforest and woodland with dominant Araucaria spp., in the drier southern part of the study area Notophyll vine forest in which many tree crowns become sparse in the dry season, true deciduous species generally absent


Detailed descriptions of rainforest classes are given in Tracey (1982).

the Wet Tropics introduced additional two classes. In this analysis, we use the simplified 15 major forest types (Table 1) defined by Hilbert and van den Muyzenberg (1999) and Hilbert et al. (2001). The typing is a pragmatic approach that is largely based on the ability to identify different types from aerial photography. The region supports about 1000 tree species with a large proportion of endemics (Hyland et al., 1998; Goosem et

al., 1999), making regional climate change assessments at the species level cumbersome, if not impossible. Rather than being based on indicator species or common species assemblages, aspects of forest structure that are strongly forced by environmental conditions were used to map the region’s forests. The Webb and Tracey vegetation classification would allow an establishment of potentially very different plant communities in the same environment. However, forest structure is


B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

not independent from the floristic composition. Classification results based on species composition show a good relationship with the structural types (Williams and Tracey, 1984). Therefore, the transition of forest types also implies an associated change of plant communities and transitions between forest types therefore spatially depend on the availability of propagules in the neighbourhood.

3. The ANN approach In general, artificial neural networks can be considered empirical models that have the capa-

Fig. 2. Artificial neural network and combined model concepts. The network is trained using today’s climate, geological information and terrain data (current environmental conditions). In the combined model, the predicted environmental suitability for different forest types is combined with the spatial structure of neighbourhood (note the selection of pixels falling into a circle) and information pertaining to ecologically likely or impossible transitions.

bility to learn the relationships between patterns— in our case, the relationships between environmental conditions and the forest pattern. We used a feedforward backpropagation network with 22 input nodes, 80 hidden nodes and 15 output nodes (Hilbert and van den Muyzenberg, 1999). The 22 input nodes represent seven climatic variables (annual temperature, minimum temperature of coldest period, temperature of warmest quarter, temperature of coldest quarter, annual precipitation, precipitation of the wettest quarter, and precipitation of the driest quarter), eight binary dummy variables for nine soil parent material classes, and seven topographic variables (slope, soil water index, slope aspect in north-east and south-west directions, respectively, and the distance to the coast, to drainage lines and perennial streams). The output of the ANN model is an environmental suitability index for any of the 15 forest types (Hilbert and van den Muyzenberg, 1999). The type with the highest environmental suitability is the most likely outcome for any location. Hence, if supplied with slightly changed maps of input variables (i.e. environmental conditions after climate change), the best-suited forest type under future conditions can be estimated (Fig. 2). The ANN model was trained using 75 000 randomly selected pixels. A further 45 000 locations were randomly selected to fit the model. The model was tested using the entire set of 1 724 893 data points. The 15 types could be discriminated with a pixel/pixel accuracy of 74.7%. The remarkably good results show a strong relationship between environmental conditions and vegetation types. The classification includes seral types (VFAE, Table 1) that are dependent on disturbances. The ANN basic model shows the lowest accuracy for these types (only 28% correct)— hence correctly identifying VFAE as co-occurring in the same environments with the late successional type mesophyll vine forests. Thirty-five percent of the VFAE current distribution was erroneously mapped as MVF (see Hilbert and van den Muyzenberg, 1999 for details). These types reflect the situation that the same environmental conditions can support different vegetation types.

B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

4. Spatial and ecological constraints: the cellular automata approach Forced by environmental changes, the transition of a forest patch to a different type is not only determined by the environmental suitability as predicted from the ANN, but also by the forest types in the proximity of the location providing propagules to change the community structure. The cellular automata approach incorporates an abstraction of the spatial influence. In the model (Fig. 2), a type can only change if the new type occurs in the proximity. We allow the neighbourhood under consideration to vary in size and test the effect of using a distance-weighted approach to explore a wide range of possible dispersal intensities (see Eqs. (1)–(3)). In addition to the spatial constraints, we incorporate the means to consider ecological constraints into the model structure by prohibiting unrealistic transitions based on ecological field knowledge (see later). In the cellular automata model, for every iteration k, a probability of change Pi “ j (k) from the current vegetation type i to a new type j is computed at any location as Pi “ j (k)=ESj ·Mi, j ·Nj (k − 1)


with ESj as the environmental suitability for type j for the changed environment (ANN output), and M a transition matrix (explained later) that is initially set to unity for all elements. Nj (k − 1) represents the sources of propagules in the neighbourhood of the location in the previous iteration. The most common cellular automata neighbourhood is the Moore neighbourhood— a 3 ×3 block including the location of interest itself. However, plant dispersal might not be limited to such a small neighbourhood. We have therefore tested the effect of different neighbourhood sizes on the predictions and furthermore evaluated the effect of using a distance-weighted number of neighbours. In all cases, the neighbourhood shape is approximately circular, defined as the set of pixels whose midpoints lie within the circle defined by a given radius (Fig. 2). The location itself is included in the computations. In the first case, Nj (k − 1) is set to 1 if one or more neighbours of the type j are found in the


neighbourhood, and 0 otherwise. In the second case, the neighbourhood (n neighbours) is evaluated using a weighted distance approach: n

% w(n)·fj (n) Nj (k−1)=




% f(n) i=1

with fj (n) being 1 if the vegetation of a neighbour is of type j, and 0 otherwise. The weights w(n) change linearly with the distance d(n) from unity at the window midpoint to 0 at the radius dmax with

w(n)=max 0,1−

d(n) dmax



The type with the highest probability is predicted for the current iteration. We have only incorporated a deterministic prediction but, as is evident from Eq. (1), a stochastic approach could be easily implemented if desired. The iteration is continued until the prediction reaches a stationary state as an objective and consistent criterion. Since we do not have information about the velocity of change, the best way to objectively address spatial constraints is by adopting this limiting case and assessing locations that are never reached by a new type. For computational efficiency when using more than nine neighbours, a proportion of pixels is skipped. This has the consequence that an equilibrium is not reached in several instances and the final maps exhibit cyclical changes of up to 30 pixels. However, considering the total of 1.7 million forested pixels of the entire region such differences are negligible.

5. Ecological constraints In addition to environmental conditions alone, a transition of a vegetation type might also depend on internal vegetation dynamics. An example would be successional seres that only exist temporarily after disturbance and have a preferred direction of change. Cyclones and fire are important factors shaping the region’s vegetation pattern, and some types (e.g. VFAE) are indica-


B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

Fig. 3. Biological interaction matrices showing the possibility of transition from type i (rows) to type j (columns). Only the possible transitions are shown. The remaining elements of the matrices are set to 0.

tive of past disturbances. Transitions may be impossible directly or show hysteresis effects. For example, whereas the change of sclerophyll forest to rainforest can be gradual (invasion of rainforest species), the change in the opposite direction only occurs abruptly (fire). Such ecological constraints motivated us to incorporate a transition matrix into the model. The matrix M (Eq. (1)) in its simples form (only 0 or 1) gives us the means to exclude directions of change. To assess ecological constraints, we procured information from selected spatial gradients and different expert opinions of the likelihood of internal vegetation type transitions. The first set of transition rules (matrix ‘env’; Fig. 3a) was generated from a summary of observed adjacency of different forest types since the spatial proximity may best reflect likely future transitions. During the compilation of the vegetation map of the region, Tracey (1982) assessed selected environmental gradients. Since observed vegetation boundaries may not reflect the entire spectrum of possible transitions, the matrix ‘env’ can be seen as the most restrictive case. In addition, we have consulted two scientists with long field experience in the region about the possibilities of change. In the matrix generated by Keith Sanderson, all transitions were set to 1 that appear possible given

enough time and a large enough environmental change (matrix ‘many’; Fig. 3b). This matrix, for example, excludes the possibility that MVF can develop in environments for ‘Mountain Rock Pavements’. Graham Harrington generated a matrix of transitions focusing on successional states (matrix ‘some’; Fig. 3c). Here, only the possible direct transitions between types are set to 1. For example, in the matrix ‘some’, MLW first changes to MOFW before it can turn into TOFW if moisture availability is increased.

6. Results

6.1. Neighbourhood constraints The regional scenario (increase of mean annual temperature by 1 °C and a decrease of mean annual precipitation by approximately 10%) is evenly applied throughout the area. The resulting map using the ANN model for this scenario predicts changes in forest type at 37% of the total forested area. These are the prediction results making the POCEC assumption against which the spatially dynamic model results are compared. The combined ANN/Cellular Automata/Transition model allows a relaxation of the POCEC

B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

assumption and ecological constraints separately. In the combined model, we use the original vegetation map as the initial condition and the ANN predictions of the climate change scenario as the final conditions that the model would approach if there were no spatial or ecological constraints. The effect of changing assumptions is assessed by comparing the ANN predictions with the spatially dynamic model results. In the following, we present the proportions of the region’s forested area that differ between the predicted maps, respectively. This percentage (comparing predicted maps) best represents the effect of model modifications. It is slightly different from the decrease in the area that is predicted to change. In the first model series, we examine how the spatial arrangement of the vegetation mosaic alone might restrict the potential change (Fig. 4a). The most restrictive case is the distance-weighted Moore neighbourhood, resulting in a difference of 10% in the area. Even if a location becomes more suitable for a different forest, a transition is not


allowed if the number of neighbours for another, less suitable type outweighs the best suited type. If the presence/absence rule is imposed, the maps only differ by up to 5%. The effect drops further if the neighbourhood size is increased. The resulting difference between the maps using the distance-weighted approach does not drop below 5%. An increase of the neighbourhood size reduces the spatial constraints and consequently results in maps that are more similar, but the overall effect of spatial constraints levels out at a radius around 2.5 km.

6.2. Ecological constraints In the second test series (Fig. 4b), we additionally consider constraints due to ecological limitations on change. Unlike the first model series where we allowed transitions between all types (all Mi, j from Eq. (1) are set to 1), we here employ the matrices ‘many’, ‘some’, and ‘env’ (Fig. 3). The resulting maps differ substantially from the ANN predictions. Corresponding to the smallest number of restrictions, the matrix ‘many’ has the lowest effect. Most surprising was the strong effect of matrix ‘some’ in restricting the changes to the equilibrium between vegetation and the environment at most locations. This effect is relatively independent of the neighbourhood size. Also, the use of distance-weighted versus presence/absence has relatively little effect.

6.3. Relati6e changes of indi6idual forest classes

Fig. 4. Percent difference between the predictions using the ANN alone and the combined approach employing the cellular automata (a) and, in addition, the ecological transition matrices (for details, see text).

The relative change of individual types under different model treatments allows us to evaluate our certainty of how different forests may be affected. Fig. 5 shows the change of the total extent of the five most extensive types for a selected set of models. We compare the four different transition rules and a neighbourhood size of 0.1 and 2.5 km radius using the distance-weighted approach. The results show substantial differences at the detail of individual types. The predicted changes of MVF and TOFTW differ strongly. In contrast, the predicted changes of CNVF, SNSM and MLW are relatively consistent. Similar to the lumped results, the response of individual types is

B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224


Fig. 5. Change (% of study area) of the five most frequent forest types in the region showing the variability among model runs.

more influenced by restrictions from the transition matrix than the pure spatial constraints. Using this selected set of model runs, we can compute the variation between the different models to quantify the uncertainty of predicting the potential response to climate change (Table 2). Expressed as the coefficient of variation (variance/mean ratio), the results are again shown to be relatively consistent for CNVF, SNSM, and MLW, whereas the extent of TOFTW and MVF varies substantially.

Table 2 Average climate change response of major forest types of different models to an increase of 1 °C and a decrease of 10% precipitation Type

Mesophyll vine forests Complex notophyll vine forests Simple notophyll and simple microphyll Tall open forests and tall woodland Medium and low woodlands

Average change (km2)

Coefficient of variation (%)











7. Discussion and conclusions Substantial differences in the responses to climate change can be observed if the POCEC assumption of the environment–vegetation relationships is relaxed, even when using a moderate climate change scenario. The results show that tropical forest types in the ‘Wet Tropics’ of Northern Australia are sensitive to assumptions related to spatial constraints, especially if the spatial context is combined with restrictions on ecological transitions. Spatial constraints alone, however, account for relatively small effects. This is an important conclusion for the region’s potential loss of biodiversity. Spatial extent and connectivity is considered a likely reason for regional species survival during climate change (Colinvaux et al., 2000). The results show that the region’s rainforest is well connected and that such a moderate climate change may allow vegetation to move to environments that are best suitable. Small changes of environmental conditions may well be absorbed by the ability of forest types to spread. Even when using the most limiting neighbourhood constraints for the cellular automata model (the 3× 3 ‘Moore’ environment), the model predictions are similar to the results from the ANN. A neighbourhood size larger than 2.5 km effectively yields the same results as the neural net alone. In other words, if a dispersal distance of 2.5 km was

B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

possible for most species, the region would be relatively insensitive to spatial constraints. A simple way to incorporate a surrogate for species competition is to link transitions with the number of propagules in the neighbourhood. This is achieved by replacing the simple presence/absence rule with a distance-weighted approach. Types occurring with a higher frequency can persist in locations where environmental conditions would be more advantageous for others. If this process is considered, a transition is hindered for a substantially higher proportion of pixels. Yet, the overall effect of spatial constraints remains low. This has important implications for the area where most species are limited in their dispersal. Of 1166 tree species contained in the electronic identification key for Northern Australia (Hyland et al., 1998), only 147 are obviously adapted to wind dispersal, i.e. have either winged or plumed fruit or seeds. A total of 562 species have dry fruit and most of the tree species have fleshy fruit— an indication of animal and/or gravitational dispersal. Large frugivores are common in the area and a seed shadow on the order of a few hundred meters can be readily observed (for example, Janzen et al., 1976; Westcott and Graham, 2000). Our results indicate that the pure spatial constraints are negligible for dispersal distances of less than 2.5 km. Fruit bats or cassowaries can potentially transport seeds several kilometres, a capability that may play an important role for forests to reach their preferred environments. The topographic conditions in the area favour the shift of vegetation boundaries. Climate gradients are extremely steep with temperature and precipitation ranges of 10 °C and 7000 mm within less than a 100 km horizontal distance, respectively, hence POCEC-based boundaries do not need to move large horizontal distances in much of the region. In addition, the location of mountain ranges roughly parallels the coastline with similar longitudinal vegetation gradients throughout the region, which in concert reduces spatial constraints. The high connectivity that is shown by our results supports a biogeographical argument for the region’s exceptionally high biodiversity. The orographic situation may have pro-


vided well-spaced refugia during glaciation periods (see Hopkins et al., 1993, 1996)— a likely reason that this is one of most diverse regions of Australia. Large differences between the predictions from the ANN and combined automata models can be observed for all transition matrices including expert opinions on ecological constraints. We had expected a large difference using the matrix ‘env’ because this matrix only contains a limited set of transitions and, in fact, one reason for its selection was to set an extreme case. But the strong effects of the other two matrices for both the distance-weighted approach and the presence/absence neighbourhood approach was surprising. It illustrates the importance of patch-level ecological processes— a scale where data are very difficult to acquire considering the regional variability. Intrinsic ecological processes, here simplified by relatively subjective and diverse expert opinion, may be the strongest limit to our predictive understanding of the region’s sensitivity to change. The predicted change in the extent of individual forest types shows a wide spectrum of possible responses with differing model structures. Environmental changes will clearly affect those types most strongly that are found in the extreme ranges of the environmental variables (Hilbert and Ostendorf, 2001). Hence, these types tend to either increase or decrease whereas types found in the medium ranges might not be strongly affected. This explains the general tendency of the climate change scenario. The warmer and drier conditions favour MLW and tend to decrease CSVF and SNSM on the colder and wetter environmental conditions. MVF and TOFTW, which form bands in an intermediate climate range, are shifted with relatively little net effect. TOFTW deserves special attention as a habitat for a variety of endangered species (Laurance, 1997; Williams and Marsh, 1998). Our results show that the delicate balance between increase and decrease of the area depends on the different ecological assumptions. The effect of decreasing the spatial constraints (increasing the influence radius) is negligible, causing small changes in either positive or negative directions. It appears enticing to interpret model iterations as seres. This, however, is probably an over-inter-


B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

pretation of the model. Iterations cannot be associated with time because field observations for regional dispersal distances are impossible. Dispersal rates and vectors can by definition be only assessed at the species level, but to achieve a predictive understanding is an impossible task due to the high biodiversity of the region. Even for much simpler systems such as boreal forest, these ecological base data are hard to come by because of the stochasticity of the dispersal process itself (for example, Clark et al., 1998). Rare and single events such as cyclones or occasional long-distance animal dispersal may be a most important long-term long-range dispersal mechanism that is entirely inaccessible to experimentation. Yet, one observation from the general behaviour of the model dynamics seems noteworthy. The relative change of individual forest types in the first iteration is unrelated to the final results. Rather, the direction of the initial change is repeatedly the reverse of the following model steps. Types that expand during the first iteration are often the long-term losers. Their ability to invade is purely a factor of lack of availability of propagules for the best-suited type. This may indicate that current observations of short term (e.g. decades) vegetation change may be transitory and not directly indicate long-term vegetation dynamics. The regional effect of climate change is to a large degree determined by the inertia of vegetation spread, a function of both velocity of propagation and the spatial arrangement of sources of propagules. This paper deals exclusively with the second aspect. However, in spite of several limitations of the approach (subjective typing, coarse resolution, etc.), we still believe that the methodology represents a valuable improvement over models based on the POCEC assumption. It allows the magnitude of a common type of model error to be assessed and therefore increases confidence in the model results. The cellular automata algorithm is computationally intense but relatively simple to implement. It is not limited to artificial neural networks and can be linked with any vegetation model quantifying vegetation– environment relationships and is therefore widely applicable. The use of simple transition matrices is a further enhancement to employ qualitative information.

The modelling framework provides the means to test a range of concepts and to evaluate the consequences of different perceptions of individual researchers or, as in our case, different qualitative information. Using a set of model structures furthermore allows identifying the relative certainty of prediction details.

Acknowledgements The authors thank Brett Buckley for programming assistance. Keith Sanderson, Graham Harrington and Andrew Graham assisted in generating the transition matrices. Vegetation, geology, and digital elevation data were supplied by the Wet Tropics Management Authority. Arnon Accad, Andrew Graham, Janet Franklin, and two anonymous reviewers provided many valuable comments to the manuscript.

References Austin, M.P., 1998. An ecological perspective on biodiversity investigations: examples from Australian eucalypt forests. Annals of the Missouri Botanical Garden 85, 2 – 17. Bazzaz, F.A., 1998. Tropical forests in a future climate, changes in biological diversity and impact on the global carbon cycle. Clim. Change 39, 317 – 336. Bolliger, J., Kienast, F., Bugmann, H., 2000. Comparing models for tree distributions, concept, structures, and behavior. Ecol. Model. 135, 89 – 102. Brzeziecki, B., Kienast, F., Wildi, O., 1995. Modeling potential impacts of climate-change on the spatial-distribution of zonal forest communities in Switzerland. J. Veg. Sci. 6 (2), 257 – 268. Bugmann, H.K.M., 1996. A simplified forest model to study species composition along climate gradients. Ecology 77, 2055 – 2074. Carey, P.D., 1996. DISPERSE: a cellular automaton for predicting the distribution of species in a changed climate. Global Ecol. Biogeogr. 5 (4 – 5), 217 – 226. Chave, J., 1999. Study of structural, successional and spatial patterns in tropical rain forests using TROLL, a spatially explicit forest model. Ecol. Model. 124, 233 – 254. Clark, J.S., Fastie, C., Hurtt, G., Jackson, S.T., Johnson, C., King, G.A., Lewis, M., Lynch, J., Pacala, S., Prentice, C., Schupp, E.W., Webb, T., Wyckoff, P., 1998. Reid’s paradox of rapid plant migration —dispersal theory and interpretation of paleoecological records. BioScience 48 (1), 13 – 24.

B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224 Colinvaux, P.A., De Oliveira, P.E., Bush, M.B., 2000. Amazonian and neotropical plant communities on glacial time-scales: the failure of the aridity and refuge hypotheses. Quaternary Sci. Rev. 19 (1 –5), 141– 169. Emanuel, W.R., Shugart, H.H., Stewenson, M.P., 1985. Climatic change and the broad-scale distribution of terrestrial ecosystem complexes. Clim. Change 7, 29 –43. Franklin, J., 1995. Predictive vegetation mapping: geographic modelling of biospatial patterns in relation to environmental gradients. Prog. Phys. Geog. 19 (4), 474 –499. Franklin, J., 1998. Predicting the distribution of shrub species in southern California from climate and terrainderived variables. J. Veg. Sci. 9 (5), 733 – 748. Goosem, S., Morgan, G., Kemp, J.E., 1999. Wet tropics. In: Sattler, P., Williams, R. (Eds.), The Conservation Status of Queensland’s Bioregional Ecosystems. Environmental Protection Agency, Queensland Government, Brisbane, pp. 7/1– 7/73. Gottfried, M., Pauli, H., Grabherr, G., 1998. Prediction of vegetation patterns at the limits of plant life: a new view of the alpine-nival ecotone. Arctic Alpine Res. 30 (3), 207 – 221. Guisan, A., Zimmermann, N.E., 2000. Predictive habitat distribution models in ecology. Ecol. Model. 135, 147 –186. Guisan, A., Theurillat, J.P., Kienast, F., 1998. Predicting the potential distribution of plant species in an Alpine environment. J. Veg. Sci. 9 (1), 65 –74. Hilbert, D.W., Ostendorf, B., 2001. The utility of artificial neural networks for modelling the distribution of vegetation in past, present and future climates. Ecol. Model. (in press). Hilbert, D.W., van den Muyzenberg, J., 1999. Using an artificial neural network to characterise the relative suitability of environments for forest types in a complex tropical vegetation mosaic. Diversity Distributions 5, 263 – 274. Hilbert, D.W., Ostendorf, B., Hopkins, M., 2001. Sensitivity of tropical forests to climate change in the humid tropics of North Queensland. Aust. Ecol. (in press). Holdridge, L.R., 1947. Determination of world plant formations from simple climatic data. Science 105, 367 – 368. Hopkins, M.S., Ash, J.E., Graham, A.W., Head, J., Hewett, R.K., 1993. Charcoal evidence for the spatial extent of the Eucalyptus woodland expansions and rainforest contractions in North Queensland during the late Pleistocene. J. Biogeogr. 20, 357 –372. Hopkins, M.S., Head, J., Ash, J.E., Hewett, R.K., Graham, A.W., 1996. Evidence of a Holocene and continuing recent expansion of lowland rainforest in the humid, tropical North Queensland. J. Biogeogr. 23, 737 –745. Huntley, B., Berry, P.M., Cramer, W., McDonald, A.P., 1995. Modelling present and potential future ranges of some European higher plants using climate response surfaces. J. Biogeogr. 22 (6), 967 –1001. Hyland, B.P.M., Whiffin, T., Christophel, D.C., Gray, B., Elick, R.W., Ford, A.J., 1998. Australian tropical rain


forest trees and shrubs. CSIRO Publishing, Colllingwood, 95 pp. Janzen, D.H., Miller, G.A., Hackforth-Jones, J., Pond, C.M., Hooper, K., Janos, D.P., 1976. Two Costa Rican bat-generated seed shadows of Andira inermis (Leguminosae). Ecology 57 (5), 1068 – 1075. Jeltsch, F., Wissel, C., 1994. Modeling dieback phenomena in natural forests. Ecol. Model. 75, 111 – 121. Kickert, R.N., Tonella, G., Simonov, A., Krupa, S.V., 1999. Predictive modeling of effects under global change. Environ. Pollut. 100 (1 – 3), 87 – 132. Kienast, F., Wildi, O., Brzeziecki, B., 1998. Potential impacts of climate change on species richness in mountain forests— an ecological risk assessment. Biol. Conserv. 83 (3), 291 – 305. Laurance, W.F., 1997. A distributional survey and habitat model for the endangered northern Bettong Bettongia tropica in tropical Queensland. Biol. Conserv. 82 (1), 47 – 60. Leathwick, J.R., 1998. Are New Zealand’s Nothofagus species in equilibrium with their environment? J. Veg. Sci. 9 (5), 719 – 732. Loehle, C., LeBlanc, D., 1996. Model-based assessments of climate change effects on forests: a critical review. Ecol. Model. 90 (1), 1 – 31. McMahon, J.P., Hutchinson, M.F., Nix, H.A., Ord, K.D., 1995. ANUCLIM User’s Guide, version 1. Centre for Resource and Environmental Studies, Australian National University, Canberra, 90 pp. Ostendorf, B., 1995. Ra¨ umlich differenzierte Modellierung von O8 kosystemen. Bayreuther Forum O8 kologie Band 13, BITO8 K, Bayreuth, 141 pp. Ostendorf, B., 1996. The influence of hydrological processes on spatial and temporal pattern of CO2 soil efflux from an arctic tundra catchment. Arctic Alpine Res. 28 (3), 316 – 325. Ostendorf, B., Reynolds, J.F., 1993. Relationships between a terrain-based hydrologic model and patch-scale vegetation pattern in an arctic tundra landscape. Landscape Ecol. 8 (4), 229 – 237. Ostendorf, B., Reynolds, J.F., 1998. A model of arctic tundra vegetation derived from topographic gradients. Landscape Ecol. 13, 187 – 201. Ostendorf, B., Quinn, P., Beven, K., Tenhunen, J.D., 1996. Hydrological controls on ecosystem gas exchange in an arctic landscape. In: Reynolds, J.F., Tenhunen, J.D. (Eds.), Landscape Function and Disturbance in Arctic Tundra. In: Ecological Studies Series, vol. 120. Springer, Berlin, pp. 369 – 386. Peng, C., 2000. From static biogeographical model to dynamic global vegetation model: a global perspective on modelling vegetation dynamics. Ecol. Model. 135, 33 – 54. Shugart, H.H., 1984. A Theory of Forest Dynamics: The Ecological Implications of Forest Succession Models. Springer, New York, 278 pp.


B. Ostendorf et al. / Ecological Modelling 145 (2001) 211–224

Shugart, H.H., Hopkins, M.S., Burgess, I.P., Mortlock, A.T., 1980. The development of a succession model for subtropical rain forest and its application to assess the effects of timber harvest at Wiangaree State Forest, New South Wales. J. Environ. Manage. 11, 243 –265. Tracey, J.G., 1982. The Vegetation of the Humid Tropical Region of North Queensland. CSIRO, Melbourne, 124 pp. Tracey, J.G., Webb, L.J., 1975. Vegetation of the Humid Tropical Region of North Queensland (15 maps at 1:100 000 scale +key.) CSIRO Australia Long Pocket Laboratories, Indooroopilly, Qld. Walsh, K., Hennessy, K., Jones, R., Pittock, B., Rotstayn, L., Suppial, R., Whetton, P., 2000. Climate Change in Queensland under Enhanced Greenhouse Conditions. CSIRO Atmospheric Research, Aspendale, 129 pp. Webb, L.J., 1959. A physiognomic classification of Australian rain forests. J. Ecol. 47, 551 – 570. Webb, L.J., 1968. Environmental relationships of the structural types of Australian rain forest vegetation. Ecology 49, 296 – 311.

Webb, L.J., 1978. A general classification of Australian forests. Aust. Plants 9, 349 – 363. West, D.C., Shugart, H.H., Botkin, D.B., 1981. Forest succession: concepts and application. Springer-Verlag, New York, 517 pp. Westcott, D.A., Graham, D.L., 2000. Patterns of movement and seed dispersal of a tropical frugivore. Oecologia 122, 249 – 257. Wiegand, T., Moloney, K.A., Milton, S.J., 1998. Population dynamics, disturbance, and pattern evolution: Identifying the fundamental scales of organization in a model ecosystem. Am. Nat. 152 (3), 321 – 337. Williams, S.E., Marsh, H., 1998. Changes in small mammal assemblage structure across a rain forest open forest ecotone. J. Trop. Ecol. 14, 187 – 198. Williams, W.T., Tracey, J.G., 1984. Network analysis of northern Queensland tropical rainforests. Aust. J. Bot. 32, 109 – 116. With, K.A., King, A.W., 1998. Dispersal success on fractal landscapes: a consequence of lacunarity thresholds. Landscape Ecol. 14 (1), 73 – 82.