The roles of statistical inference and historical sources in understanding landscape change: the case of feral buffalo in the freshwater floodplains of Kakadu National Park

Arguably the greatest challenge for the discipline of biogeography is to develop an understanding of the temporal trajectories of the spatial patterns of life on Earth. Global environmental change, concurrent debates about appropriate conservation interventions in response to the biodiversity crisis, and sustainable modes of land management hinge on understanding the trajectories of species populations and the quality of habitats. The case of the exotic water buffalo [Bubalus bubalis (Linnaeus)] populations in Kakadu National Park (KNP) is a prime example of the importance of historical data in understanding landscape change.

The KNP has a complex history. In the 200 years since European settlement this region has transited from a collection of hunter–gatherer estates, with roots in the late Pleistocene, to become Australia’s largest national park (Mulvaney & Kamminga, 1999). Before the declaration of the park, the region sustained a buffalo hide industry, and had localized pastoral leases, mines and tourism ventures (Press et al., 1995). Most recently, Kakadu has earned the dubious reputation as an iconic Australian natural environment directly threatened by rising sea levels associated with global climate change (Hennessy et al., 2007). There is no question that an understanding of the trajectory of landscape change within KNP is important. The problem is how to discover the drivers of changes. To this end a large research programme was established to gain an understanding of landscape change in the park, which was funded by both the KNP and the Australian Research Council (see Walden & Nou, 2008). The aim was to assemble historical data, derived primarily from historical aerial photography, for statistical analyses. The programme has yielded novel insights into landscape change, the most striking of which is the park-wide trend of increasing woody biomass (Banfai & Bowman, 2006, 2007; Bowman & Dingle, 2006; Banfai et al., 2007; Bowman et al., 2008; Lehmann et al., 2009).

There are a number of challenges associated with understanding spatio-temporal trends from the aerial photographic record. These include: (1) the vexed statistical issue of spatial and temporal autocorrelation; (2) the difficulty in interpreting the episodic spatial coverage of photography, namely that although it provides highly detailed spatial data it gives but mere ‘snapshots’ of what is a continuously variable system; and (3) the difficulty in sourcing spatio-temporal environmental data, comparable in terms of scale and frequency to the idiosyncratic aerial photographic record, that can be used as predictor variables in statistical modelling (Brook & Bowman, 2006; Bowman et al., in press). Bowman et al. (2008) sought to address these concerns in their study of the impact of feral buffalo on the lowlands of KNP, which was based on the analysis of historical aerial photographs, by (1) developing a robust index of feral buffalo populations, (2) using a natural experiment provided by an existing buffalo enclosure, (3) assessing the effect of surrounding vegetation, and (4) following an analytical pathway to control for spatial autocorrelation. The buffalo density index was based on counting the number of buffalo tracks that intersected a grid placed over the aerial photographs. This approach provided the best high-spatial-resolution index of buffalo densities in KNP. Satellite fire mapping was not used in the analysis because of the mismatch of the scale of the aerial photography (pixel width of 30 vs. 1 m respectively) and because it was available for only half of the 40-year study period. Nonetheless, it is widely assumed that fire activity on the floodplains has increased since the control of feral buffalo in the 1980s (Russell-Smith et al., 1997; Gill et al., 2000).

We contend that the findings of Bowman et al. (2008) have been misunderstood by Petty & Werner (2009). These authors had previously advanced a hypothesis that buffalo were the cause of dramatic changes to KNP lowlands (Petty et al., 2007). The analysis of Bowman et al. (2008) did not support Petty et al.’s (2007) hypothesis. The hypothesis of Petty et al. (2007) was based on a synthesis of historical information and on the preliminary findings of woody cover change derived from three student projects (Riley, 2005; Banfai, 2007; Lehmann, 2009). These student projects formed part of the KNP landscape change project and have now been published (Banfai & Bowman, 2006, 2007; Banfai et al., 2007; Bowman et al., 2008; Lehmann et al., 2009). Here it is important to note that because Petty et al. (2007) used the preliminary findings of these student projects, in both Petty et al. (2007) and Bowman et al. (2008) the data were derived from the same historical aerial photographs and using the same techniques: the point of difference hinges on (1) the assignment of buffalo densities, (2) the use of advanced statistics versus elementary statistics, and (3) the interpretation of the trends apparent in the data.

To illuminate some issues in the analysis of landscape change associated with Petty and Werner’s concerns we present and interpret the raw data of woody expansion on treeless floodplains that underpinned the recent analysis by Bowman et al. (2008). We focus on the floodplains rather than on savannas because from aerial photography any buffalo effect would be most obvious in driving the transition from non-woody to woody vegetation, rather than on the thickening of existing woody vegetation. Our mapping of woody vegetation from the historical sequence of five aerial photography coverages between 1964 and 2004 shows that woody cover has increased from 1964 onwards on the freshwater floodplains of KNP. There has been an abrupt increase in woody vegetation cover since 1984, and particularly since 1991 (1, 2). Indeed, the trend of woody thickening is ongoing, as is apparent when our mapped data from 2004 are overlain on the most recent Google Earth imagery (see Fig. S1 in the Supporting Information). A conspicuous feature of the woody vegetation expansion is that it occurs incrementally from existing woody patch boundaries rather than across the landscape. The previous statistical analysis showed that the vegetation ‘edge effect’ explained 24.6% of the deviance in the data describing the change of samples from treeless to woody vegetation (Bowman et al., 2008).

The detailed mapping of buffalo tracks for the study areas for the five aerial photographic coverages reflects the striking collapse of buffalo populations between 1975 and 1984 as a result of the park-wide buffalo control programme (Fig. 2) that ran between 1980 and 1992 (Robinson & Whitehead, 2003). However, the feral animal control methods resulted in considerable spatial variability in buffalo density over the park (Robinson & Whitehead, 2003), so that localized areas with high track density remained in 1991 and 2004 (Fig. 2). Bowman et al. (2008) found that the inclusion of surrounding buffalo track density as a continuous variable in the statistical model of woody vegetation encroachment improved the explanatory power by only 0.35%. This small effect was nonetheless included in Bowman et al.’s (2008) predictions of vegetation change through the use of weighted model averaging. The assertion by Petty and Werner that Bowman et al. (2008)‘mislabelled’ buffalo densities hinges on the assumption that the various historical records assembled by Petty et al. (2007) are better than our index of buffalo density: we reject that assumption.

A superficial inspection of the maps in Fig. 2 invites the conclusion of a causal relationship between buffalo numbers and woody expansion through time. In this context, it is interesting to note that the first Kakadu management plan, written in 1980, stated that it was ‘impossible to quantify’ the relative impacts of buffalo on the park’s environment (Robinson & Whitehead, 2003). For example, the management plan noted that although buffalo trampled vegetation, they reduced fuel loads and hence decreased the incidence of destructive fires. The approach taken by Bowman et al. (2008) was that proximate buffalo density is likely to be related to putative impacts on vegetation, and therefore this variable was included in the statistical models as a continuous variable. Petty et al. (2007) posit an indirect relationship between buffalo and woody vegetation encroachment, for which there appears to be no straightforward test beyond a general relationship between temporal variation in regional buffalo density and woody vegetation encroachment. This qualitative approach confounds other factors that have changed through time, blunting any inferential reliability.

There is no doubt that buffalo have a negative effect on woody cover, particularly where they change hydrological conditions and cause salinization (e.g. Stocker, 1971). Indeed, Bowman et al. (in press) discovered a relationship between high levels of buffalo (using tracks on 1964 aerial photographs) and the contraction between 1964 and 1984 and 1984 and 2004 of Melaleuca swamp forests that form a fringe around some freshwater floodplains. The mechanism of the retreat appears to be related to salinization, given that the retreat was greatest on low-lying sites vulnerable to penetration during spring tides. This finding shows that the use of localized buffalo density can be statistically related to vegetation dynamics.

Woody expansion has been noted in the savannas and monsoon rain forest patches throughout KNP and elsewhere in northern Australia (e.g. Bowman et al., 2001; Brook & Bowman, 2006). Recent work has disclosed a comparable woody expansion of mangroves into adjacent salt-flats near Darwin (Williamson et al., in review). This latter study is significant in the debate about the drivers of woody vegetation expansion in the monsoon tropics because neither fire nor buffalo disturbance is a plausible confounding factor. Clearly, moving from statistical inferences to biological mechanisms is the next step in explaining the trend in increasing woody biomass in the monsoon tropics. Such research is now underway, as outlined by Bowman et al. (in review b), through the study of (1) rain forest boundary dynamics using aerial photography across a continental gradient and (2) patterns of tree growth using the native conifer Callitris as a model system.

Petty & Werner (2009) disagree with the statistical analysis and interpretations by Bowman et al. (2008), preferring to rely on selected evidence that is arranged in a particular way so as to create a narrative that buffalo densities are causing changes in the KNP landscape. Unlike Bowman et al. (2008) they have not grappled with the limitations of the various data sources, nor tested alternative hypotheses by applying statistical analyses appropriate for spatio-temporal data. Rather, they have privileged an amalgam of historical sources, previous descriptive reviews and personal testimonies, particularly sketchy historical records of buffalo densities, for example see figure 4 in Petty et al. (2007), that are of very low spatial resolution compared with the index of buffalo tracks used by Bowman et al. (2008). The use of such qualitative information to bolster narratives about landscape change has a rich tradition in the discipline of environmental history, which is founded in the humanities. Bowman (2001) has argued that although such environmental history narratives serve an important purpose in formulating hypotheses, communicating knowledge and stimulating debate, they cannot be used as a substitute for statistical analysis. There is no doubt that Petty et al. (2007) highlighted the importance of understanding the impacts of feral buffalo on an environment that had never before supported ungulates. However, environmental narratives must change in the light of new knowledge if they are to serve any scientific or management purpose.