- Email: [email protected]
The application of local measures of spatial autocorrelation for describing pattern in north Australian landscapes
Journal of Environmental Management (2002) 64, 85–95 doi:10.1006/jema.2001.0523, available online at http://www.idealibrary.com on
The application of local measures of spatial autocorrelation for describing pattern in north Australian landscapes Diane M. Pearson School of Biological, Chemical and Environmental Science, Faculty of Science Information Technology and Education, Northern Territory University, Darwin, NT 0909, Australia Received 26 October 2000; accepted 2 October 2001
This paper tests the use of a spatial analysis technique, based on the calculation of local spatial autocorrelation, as a possible approach for modelling and quantifying structure in northern Australian savanna landscapes. Unlike many landscapes in the world, northern Australian savanna landscapes appear on the surface to be intact. They have not experienced the same large-scale land clearance and intensive land management as other landscapes across Australia. Despite this, natural resource managers are beginning to notice that processes are breaking down and declines in species are becoming more evident. With future declines of species looking more imminent it is particularly important that models are available that can help to assess landscape health, and quantify any structural change that takes place. GIS and landscape ecology provide a useful way of describing landscapes both spatially and temporally and have proved to be particularly useful for understanding vegetation structure or pattern in landscapes across the world. There are many measures that examine spatial structure in the landscape and most of these are now available in a GIS environment (e.g. FRAGSTATSŁ ARC, r.le, and Patch Analyst). All these methods depend on a landscape described in terms of patches, corridors and matrix. However, since landscapes in northern Australia appear to be relatively intact they tend to exist as surfaces of continuous variation rather than in clearly defined homogeneous units. As a result they cannot be easily described using entity-based models requiring patches and other essentially cartographic approaches. This means that more appropriate methods need to be developed and explored. The approach examined in this paper enables clustering and local pattern in the data to be identified and forms a generic method for conceptualising the landscape structure where patches are not obvious and where boundaries between landscape features are difficult to determine. Two sites are examined using this approach. They have been exposed to different degrees of disturbance by fire and grazing. The results show that savanna landscapes are very complex and that even where there is a high degree of disturbance the landscape is still relatively heterogeneous. This means that treating savanna landscapes as being made up of homogeneous units can limit analysis of pattern, as it can over simplify the structure present, and that methods such as the autocorrelation approach are useful tools for quantifying the variable nature of these landscapes. 2002 Academic Press
Keywords: Landscape structure, spatial analysis, landscape ecology, continuous variation, spatial autocorrelation, northern Australian savanna landscapes.
Introduction Although the tropical savanna landscapes of northern Australia appear to be intact they are under considerable pressure from anthropogenic activity. The result is subtle alterations to the landscape
Email: [email protected] 0301–4797/02/010085C11 $35.00/0
that are resulting in declines in certain species (Price and Bowman, 1994; Braithwaite and Griffiths, 1996) and progressive changes in structure that are thought to be detrimental (Woinarksi and Whitehead, 2000). One suite of species that are affected by these changes are granivorous birds, most notably the Gouldian finch (Erythrura gouldiae), that are diminishing in numbers across northern Australia (Tidemann, 1996). 2002 Academic Press
86
D. M. Pearson
The structural modifications that these landscapes have experienced are thought to be associated with pressures from grazing and changes in fire regimes. Fire has for a long time had a dominant influence on northern Australian savanna landscapes (Pyne, 1991) and many of the native species of vegetation are adapted to cope with fire at some stage in their life cycle and are often reliant upon it. Prior to European settlement burning was practised by Aboriginal people for cultural reasons, hunting purposes, and in order to provide good grazing habitat for the wildlife they hunted (Haynes, 1985; Braithwaite and Roberts, 1995). This created a mosaic of burnt and unburnt habitat in the landscape. However, since European settlement, fire regimes have changed, and now vast tracts of the landscape are burnt on an annual basis (Russell-Smith et al., 2000). It is currently thought that recent fire regimes in parts of northern Australia have been too frequent and too expansive (particularly when held in the late dry season) to ensure adequate regeneration of some species and to maintain biodiversity (RussellSmith et al., 2000). Added to this, in recent years northern Australian savanna landscapes have been put under greater pressure from increased cattle grazing intensities (Ludwig et al., 1999). In order to achieve sustainable use of these rangelands a better understanding of their ecology, particularly their spatial and temporal structure and their functional roles, is required (Ludwig et al., 1999). Since the spatial arrangement of landscape elements and the relationship between these elements, (the complexity between and within habitat patches), and how these change over time, is thought to influence the distribution and abundance of species, it is becoming increasingly important for natural resource managers in northern Australia to be able to successfully model landscape pattern and determine the ‘healthiness’ of a landscape. A healthy landscape has been defined by the Corporate Research Centre for Sustainable Tropical Savannas as one that can (i) maintain basic landscape functionality (e.g. nutrient cycling, water capture, food and shelter for fauna) at different spatial scales, (ii) maintain viable populations of native species at appropriate spatial and temperate scales, and (iii) reliably meet the needs of people over the long term (Whitehead and Gorman, 1999). Landscape ecology together with Geographic Information Systems (GIS) technology provides a framework within which one can examine spatial and temporal differences and changes in structure and function at the landscape scale, and relate these changes to modifications in the pattern of
land use. These methods and techniques have been used extensively across the world (Stow, 1993; Bridgewater, 1993; Kienast, 1993; Walsh and Davis, 1994; McGarigal and Marks, 1994). To understand structural changes in the landscape a number of measures of landscape pattern and diversity have been devised and a comprehensive review of these indices are reported in Haines-Young and Chopping (1996). The majority of these measures are based on patch theory, which relies on the identification of the three basic landscape features – patches, matrix and corridors (Forman, 1995) by some form of land cover classification. Many of these indices are now contained within software packages which are fully integrated with GIS such as FRAGSTATSŁ ARC (Innovative GIS Solutions Inc, 1999), r.le (Baker and Cai, 1992) and Patch Analyst (Elkie et al., 1999). Most of these measures take the form of indices, which means that they describe pattern through the calculation of a single number. Most of the commonly used indices have been developed for cultural landscapes in Europe and North America where agricultural development and management over centuries has produced a richly mosaiced landscape dominated by fields, field margins, woodland remnants and other discrete types of land cover. Since traditional approaches to landscape ecology have tended to be based on the identification of patch-matrix-corridor features (Forman and Godron, 1981), in cartographic terms, this means partitioning the landscape into homogeneous units with distinct boundaries, and the geometry (area and perimeter) of these homogeneous units are used as inputs to the computation of landscape metrics. This approach works well when applied to cultural landscapes in Europe and America, and even in parts of Australia, such as the wheatbelt of Western Australia (Hobbs et al., 1993), where landscape features can be clearly identified. However, it is less relevant for modelling areas where there is a gradation of land cover and where boundaries between landscape features are not clearly identified (McIntyre and Barrett, 1992). Such gradations of land cover are evident in the tropical savannas found in northern Australia. This is also a problem encountered in other savanna landscapes in the world and in other landscapes were structural modifications are subtle, like in forested landscapes were selective logging has taken place rather than large-scale clearance (Pearson, 1998). This paper investigates the use of an alternative approach to quantifying structure and structural changes based on modelling pattern
Local measures of spatial autocorrelation
through the use of a measure of local spatial autocorrelation, integrated within a GIS (ARC/INFO (ESRI, 1998)) (Pearson, 1998). For comparison, the results of modelling landscape pattern using local autocorrelation are compared with a more traditional way of representing the pattern present in the landscape (a simple classification of land cover) that was produced by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) (Eager, pers. comm.; Pickup et al., 2000).
Study area The study sites examined in this paper are taken from the Kidman Springs Research Station which is situated in the Victoria River District (VRD) of the Northern Territory and lies about 700 km south of Darwin (see Figure 1). The Station is managed as a research pastoral property owned by the Department of Primary Industry and Fisheries and consists of rangelands/tropical savannas that are typical of the region. They are under considerable pressure from grazing and management for grazing using fire.
In the Kidman Springs area the median rainfall is approximately 640 mm, which is largely, summer (wet season) dominant (Ludwig et al., 1999). Trees and perennial grasses in the tropical north of Australia rely on being able to capture and store water and nutrients during the wet season for their growth and survival in the dry season. If plants are unable to capture and store water and nutrients they will become stressed, especially in the late dry season, which will result in increased plant mortality, particularly amongst those plants which are susceptible to grazing and fire (Williams et al., 1997; Ludwig et al., 1999). The pattern or structure of vegetation patches plays an important role in capturing and conserving important resources. Where landscapes have lost significant vegetation patches (particularly perennials) they have a reduced capacity to capture and store scarce resources. This affects the overall condition or healthiness of the landscape, which in turn influences its capacity to sustain viable populations of flora and fauna (Ludwig et al., 1999). The loss of significant vegetation patches can also have a knock on effect in terms of a loss of diversity of species of fauna that depend on these plants for their survival, and where overgrazing takes place
Darwin
Kidman Springs
Figure 1.
87
Location of Kidman Springs study site in the Victoria River District of the Northern Territory, Australia.
88
D. M. Pearson
in the rangelands, soil erosion can result, which in turn can affect the stability of the landscape and its sustainability for supporting biodiversity. Cattle tend to focus grazing and trampling near to watering points, as a result the distance from a watering point is used in this study as a surrogate for a gradient in grazing pressure (Ludwig et al., 1999). Sites were chosen so that differences in pattern could be observed relative to exposure to fire and heavy grazing.
Data The data used in this case study consist of two images of the Kidman Springs area captured using airborne videography, which have a resolution of 0Ð2 m. One of the sites depicted by an image is very close to a watering hole (KS1) the other is in an exclosure (KS5) (grazing and fire have been excluded from this area for about 25 years). The dominant vegetation in both sites is low, open eucalypt savanna with the typical grass species being perennial black spear grass (Heteropogon contortus), curly bluegrass (Dichanthium fecundum), biennial limestone grass (Enneapogon polyphyllus) and the annual, false couch (Brachyachne convergens) (Ludwig et al., 1999). Both sites are on the same land unit type. This is described as ‘gently undulating plains with calcareous red earth soils’ (Ludwig et al., 1999). Videography imagery was captured individually for the two sites for four bands (blue, green, red and infrared). A simple sequential classification procedure was employed by CSIRO using transforms and indices rather than raw bands to sequentially identify the five main land cover types in the area i.e. shadow, tree/shrub canopy, bare soil, perennial patches, and annuals/litter (Eager pers. comm.; Pickup et al., 2000). This classification is included as ancillary data in this study for comparative purposes to contrast more traditional ways of representing landscape pattern with the alternative approach put forward in this paper.
Methodology The approach examined in this study is to apply a technique for modelling pattern that was thought to be better suited for understanding the spatial heterogeneity or pattern in landscapes where cartographic description fails to capture the variability present (Pearson, 1998). This approach is
consistent with that suggested by Musick and Grover (1991) who recommend applying textural measures, which take account of spatial variation via analysis of brightness values in the data as a method of modelling spatial pattern present in landscapes. The approach explored is geostatistical and applies a local measure of spatial autocorrelation based on the statistic Geary’s C. This type of mathematical measure is referred to as a structure function (Legendre and Fortin, 1989). Measures of spatial autocorrelation (Cliff and Ord, 1973) work by examining how objects at one location are similar to objects located nearby (Goodchild, 1986), that is, they look for spatial dependence in the data (similarity as a function of distance) (Chou et al., 1990). If features situated close together have similar attribute information, then the pattern in the data can be described as exhibiting positive autocorrelation. When features close together are more dissimilar in attribute value than features further away, pattern in the data is negatively autocorrelated. Zero autocorrelation exists when attributes or their values are independent of location (Goodchild, 1986). Fieldwork shows that boundaries between patches are difficult to distinguish in this landscape with large areas of land existing as a mixture of annuals and perennials grading into one another (this is particularly the case in KS1). The grading between vegetation classes makes it difficult to clearly define boundaries around patches in these sites. This means that classifying the landscape into broad groups can over-simplify the ground conditions. As a result, in order to prevent simplifying a complex landscape the raw spectral data were used for analysis with the autocorrelation approach, thus no objects such as land cover polygons are pre-defined in the data prior to analysis. The approach has previously been applied to research conducted on New South Wales (NSW) forested landscapes (Pearson, 1998). Although the forested landscapes in NSW are different to the savanna landscapes found in northern Australia, it is thought that both types of landscape are less inclined to exhibit the clearly defined patch, matrix and corridor features usually examined in landscape ecological analysis. In contrast, they demonstrate gradations between vegetation classes across the landscape. Conventional measures of spatial autocorrelation in a GIS operate on a global basis, thus they perform simultaneous measurements of spatial dependence for all locations within the dataset and return a single value for that data set. Getis
Local measures of spatial autocorrelation
and Ord (1996) recommend the use of local statistics, which work by measuring spatial association between one pixel and its neighbours up to a certain sampling distance. An ARC/INFO AML (ESRI, 1998) program was written to produce a local measure of Geary’s C (Pearson, 1998). It operates on the principle of examining the relationship between a variable measured at a given point and the same variable measured at points some distance from the original point. This approach is similar to a method used by Fotheringham et al. (1996) for investigating spatial non-stationarity. This methodology applies the principles of landscape ecology to cells rather than patches. It enables one to look at variation not only in space and time, but also to examine the effect of changing the scale of analysis on the output, by varying the grain and extent (Turner, 1990). Since four bands of spectral information make up each videography image and the autocorrelation routine can only run on one surface at a time, a new GRID surface was created on which to perform the analysis. This was made up of the average spectral reflectance values for the four bands of data. The effect of changing the scale of analysis has an important impact on the results because all measures of landscape pattern are scale dependant (Haines-Young and Chopping, 1996), and results are meaningless unless made with reference to the scale of measurement (Turner et al., 1989). Qi and Wu (1996) looked at the effects of changing spatial resolution on the results of landscape pattern analysis using spatial autocorrelation indices. They showed that the degree of autocorrelation present in pattern is dependant on scale and that changing the scale effects the amount of autocorrelation found in pattern. Appropriate scales of analysis tend to be a function of the type of environment being examined and the sort of information that is required (Woodcock and Strahler, 1987). Craig and Labovitz (1980) and Labovitz et al. (1980) measured autocorrelation in Landsat MSS and TM images and found it to be higher in images with a fine spatial resolution than in images of much coarser resolution. Performing analysis on grids of varying size, allows the effect of changing the spatial resolution on the result to be assessed. The importance of scale is addressed in this autocorrelation measure by altering the sample size (i.e. the distance over which spatial dependence is calculated or the area that the moving window covers), and also the cell size (i.e. the resolution of analysis – this is changed using the GRID SETCELL function in ARC/INFO (ESRI, 1998)). These steps were built into the ARC/INFO
89
AML (ESRI, 1998) program so the user could run the analysis at a range of cell and window sizes and compare the results. The results are new sets of surfaces for each of the scales and sample sizes examined. Changing the scale of analysis helps to determine appropriate scales at which certain landscape features operate. In order to evaluate the impact of changing scale of analysis on the results the analysis was conducted at cell sizes starting from 0Ð2 m to 1 m and windows of analysis ranging from 1 m to 100 m. Determining the most appropriate scale of analysis is based on the user’s interpretation of the results in conjunction with patch location and size information collected in the field. In order to get some idea of the accuracy of the autocorrelation approach in terms of its ability to determine patches in the landscape, ArcView (ESRI, 1999) GIS was used to overlay the locations of patches recorded from fieldwork on top of the resulting autocorrelation surfaces. These figures were compared with overlaying the same field data on the results of the CSIRO classification for the two study sites.
Results Conducting the analysis at a range of cell sizes and windows of analysis, then examining the output overlaid with the field data in ArcView (ESRI, 1999) GIS, enabled appropriate scales of analysis to be determined for each of the study sites. The results showed that a cell size of as fine a resolution as possible was most appropriate for both sites, as this was important for maintaining detail in the image and enabling small scale pattern in the landscape to be determined. The finest resolution possible with the videography imagery used in this study is the pixel size, which is 0Ð2 m. Figure 2 shows the effect of changing the window of analysis on the output from KS1. Similar patterns were evident for KS5. Small windows of analysis produced a surface that was made up of a pattern that was very speckled in appearance and bore no resemblance to the pattern actually present in these landscapes. As the window size increased a pattern that was more consistent with vegetation patches recorded in the field emerged. Up to a certain scale, areas of positive autocorrelation start to correspond to patches recorded in the field, which could be either patches of perennials, annuals or tree canopies. However, beyond a certain window size the data structure becomes overly simplified and the influence of the edge
90
D. M. Pearson
Figure 2. The effect of changing the window of analysis on the results of autocorrelation on KSI. Window sizes (in m): (a) 4ð4; (b) 16ð16; (c) 32ð32; (d) 60ð60.
of the data becomes more evident until a point where the amount of useable data is reduced to a very narrow strip. Analysis at different scales revealed that for site KS1 a window size of 60 m ð 60 m was appropriate for modelling landscape pattern, whilst a window size of 48 m ð 48 m was deemed to be appropriate for modelling of pattern in KS5. Figures 3 and 4 show that, at the scales listed above, the autocorrelation approach produces surfaces that are able to identify some of the main patches detected by fieldwork. Patches of similar land cover type are represented as areas that are positively autocorrelated, while the surrounding areas that are more variable, are represented by regions of negative autocorrelation. What occurs is that positive autocorrelation corresponds to patches of perennials or annuals identified from fieldwork while negative autocorrelation tends to correspond to where there is a mixture of grass species (annuals and perennials grading into one another) or where there is a mix of bare ground and some grassy tussocks.
The pattern depicted in the autocorrelation surfaces seems to resemble the pattern recorded in the field more closely than the results produced from applying a simple classification to the data (see Figures 5 and 6). A comparison with the CSIRO classification data at KS1 and KS5 shows that the autocorrelation surfaces give more indication of the variability present in the landscape and therefore do not over simplify what is in effect a very complex landscape by treating the area as being capable of being defined as a series of homogeneous units. Table 1 shows that for KS1 the percentage of the area that is positively autocorrelated is 18Ð7% while 77Ð7% of the area is negatively autocorrelated. This compares with 22Ð7% and 73Ð3% respectively for KS5. These figures provide further evidence of the amount of heterogeneity present in both these landscapes. The percentage of the area that is negatively autocorrelated in KS1 is particularly a surprise if one views the classified image for this landscape. The CSIRO classification for KS1 depicts that the landscape is made up of quite large homogeneous areas; particularly the area
Local measures of spatial autocorrelation
91
Figure 3. Spatial autocorrelation surface for KSI (0Ð2 m cell size and 60ð60 m window) with location of patches recorded in the field.
Figure 4. Spatial autocorrelation surface for KS5 (0Ð2 m cell size and 48ð48 m window) with location of patches recorded in the field.
described as bare ground. The results of applying the autocorrelation approach therefore reveals that the Kidman Springs landscapes are much more variable spatially than a simple classification will allow for.
Discussion The results show that the autocorrelation approach is capable of describing more of the spatial
variability present at each of the sites examined in this paper than is evident by performing a simple classification on the same sites. This result is consistent with the idea that grassy landscapes tend to exhibit continuous variation, which make them difficult to represent in a two-dimensional surface made up of homogeneous units with distinct boundaries (Pearson and Riley, 2001). Therefore, to define patchiness or pattern in the landscape using the autocorrelation approach appears to be one way of
92
D. M. Pearson
Figure 5.
Classification of KS1 and the location of patches recorded in the field.
Figure 6.
Classification of KS5 and the location of patches recorded in the field.
addressing the problem of high spatial variability in grassy landscapes and producing surfaces that give a more realistic representation of the pattern present. The advantages of modelling pattern using spatial autocorrelation over other ways of quantifying landscape structure are that it allows environmental variation to be represented as a set of distinct descriptions at grid cells across the image and reduces the problems associated with indeterminate boundaries between landscape features. Other advantages that exist by defining landscape
Table 1. Percentage of total area for each site that is represented by each level in the autocorrelation surfaces Site
Percent positively autocorrelated
Percent zero autocorrelated
Percent negatively autocorrelated
KS1 KS5
18Ð7 22Ð7
3Ð6 3Ð6
77Ð7 73Ð7
pattern by the autocorrelation approach rather than classifying landscapes to produce surfaces made up of homogeneous regions, are that many
Local measures of spatial autocorrelation
of the fine scale patches or pattern that may be important to grassy landscapes, particularly for species like granivorous birds, may not be detected if the landscape becomes too simplified by a broad classification. For example, what is described as homogeneous ‘bare ground’ in the CSIRO model of pattern shows up to be quite spatially variable using the autocorrelation approach. Closer inspection of the field data shows that what is classified as bare ground is actually a mixture of tussocks of grass or small, close to the ground patches of grass species interspersed with bare ground. Therefore, the spectral response from remote sensing is expected to be mixed in this region so it is not appropriate to treat this area as being homogeneous if one needs to understand fine scale structure. Fieldwork also showed that KS1 has more intermixing of annuals and perennials than KS5, thus making boundaries between patches difficult to determine in the field and producing a heterogeneous effect at the fine scale, which explains the large percentage of the area that was negatively autocorrelated in the KS1 autocorrelation surface. Field investigation of the two sites showed that in comparison to KS1, KS5 had more dense patches and tended to have larger patches of mostly perennials with only small patches of annuals present. This again explains why KS5, which although appears on the surface to be more heterogeneous than KS1, actually produced autocorrelation results that show a smaller percentage of the area being negatively autocorrelated than KS1. Differences in structure observed both by fieldwork and from the autocorrelation surfaces are probably related to higher disturbance factors in KS1, in particular those caused by intensive cattle grazing and trampling close to the watering hole. KS5 has been relatively undisturbed by either fire or grazing. The fact that KS1 is seen to be more heterogeneous at fine resolution as a result of disturbance could indicate that this part of the landscape is actually more unstable than KS5, making the approach a possible indicator of landscape condition/health. The spatial autocorrelation approach is also particularly useful for showing the effect of scale on landscape studies and how different land use features operate at different scales and therefore require different scales of analysis, (KS1 required a larger window of analysis to identify the large significant patches than KS5). Patches identified by fieldwork are only found with this method by applying the analysis at a specific spatial scale, resolution or window size. The results indicate how spatial scale plays an integral role in determining the results from analysis of pattern or structure
93
present in the landscape. However, it could also be seen as being a limitation with the autocorrelation approach as, in order to ensure meaningful results, the resolution of analysis needs to be carefully selected. Another limitation of this approach is that although viewing the autocorrelation surfaces in the GIS shows that they succeeded in identifying many of the significant patches that were detected from fieldwork, there are still some discrepancies between patch locations in the field and those on the surfaces. More fieldwork is needed to be able to explain these anomalies fully and to understand why some patches are detected and others are missed by this approach. Despite some of the limitations, the autocorrelation approach has the potential to be a valuable tool for use in landscape scale studies, as it may help the user to select appropriate scales of analysis for particular landscape types. Issues of scale of analysis and detecting scales of operation of landscape features are major problems for researchers interested in landscape pattern that have not been adequately addressed up to now.
Conclusion Change to landscape structures that adversely affects their functioning can have serious implications for the health of a landscape and its biodiversity. More survey work is needed to fully understand the association between patterns of vegetation and landscape health. GIS and remote sensing technologies, coupled to modelling procedures such as the one described in this paper, can form useful surrogates for monitoring the status and condition of landscapes and therefore have the potential to be useful indicators of landscape health (Aspinall and Pearson, 2000) in the tropical savannas. The technique described in this paper is currently being applied to try to understand grassland patterns and how they relate to the sustainable management of granivorous birds in northern Australia (Pearson and Riley, 2001). Such studies involve many complexities, but the ability to successfully model some of the complex spatial associations and patterns in landscapes will ultimately help us to understand the impacts of human activity on northern Australian landscapes, and to explain recent declines of species such as the Gouldian finch from savanna landscapes. The application of geostatistical techniques such as the autocorrelation approach examined
94
D. M. Pearson
in this paper can be seen as being a positive step towards successfully modelling heterogeneous savanna landscapes. It forms a generic method of conceptualising the landscape structure with a particular scale, in areas where patch structure is not obvious and where boundaries between landscape features are difficult to determine, thus helping to deal with problems associated with indeterminate boundaries (Burrough, 1996) that are a major issue in places like northern Australia. Classification of remotely sensed imagery for grassland species is very difficult because of the complex and heterogeneous nature of the landscape. Dividing the landscape into broad categories, as usually happens in classification, enables some pattern to be identified, but the result may end up excluding pattern at a smaller scale which could in fact be the structure that is most important for species that inhabit these landscapes, such as granivorous birds (Pearson and Riley, 2001). Since the autocorrelation approach does not impose structure prior to analysis, none of the variability present in the data is simplified before analysis. As a result it may have an advantage over the currently available procedures for quantifying landscape structure, which fail to describe north Australian savanna landscapes adequately because of their highly variable nature (Pearson and Riley, 2001).
Acknowledgments I would like to thank the Northern Territory University for funding the project. CSIRO and the Corporate Research Centre for the Sustainable Development of Tropical Savannas (TS CRC) especially John Ludwig and Robert Eager, provided data and encouragement for this project. Thanks also to everyone at the Kidman Springs Research Station for their hospitality, two anonymous referees. Julian Gorman for assistance with the fieldwork and discussions on the research, and Richard Aspinall for his comments on earlier drafts of this paper.
References Aspinall, R. J. and Pearson, D. M. (2000). Integrated geographical assessment of environmental condition in water catchments: linking landscape ecology, environmental modeling and GIS. Journal of Environmental Management 59(4), 299–319. Baker, W. L. and Cai, Y. (1992). The role programs for multiscale analysis of landscape structure using the GRASS geographical information system. Landscape Ecology 7(4), 291–302.
Braithwaite, R. W. and Griffiths, A. D. (1996). The paradox of rattus tunneyi: endangerment of a native pest. Wildlife Research 23, 1–21. Braithwaite, R. W. and Roberts, S. (1995). Aboriginal burning and conservation management in Kakadu National Park, tropical Australia. Wildfire September, 29–38. Bridgewater, P. B. (1993). Landscape ecology, geographic information systems and nature conservation. In Landscape Ecology and GIS (R. Haines-Young, D. R. Green and S. H. Cousins, eds), pp. 23–36. London: Taylor and Francis Ltd. Burrough, P. A. (1996). Natural objects with indeterminate boundaries. In Geographic Objects with Indeterminate Boundaries (P. A. Burrough and A. U. Frank, eds), pp. 3–28. London: Taylor and Francis Ltd. Chou, Y. A., Minnich, R. A., Salatar, L. A., Power, J. D. and Dezzani, R. J. (1990). Spatial autocorrelation of wildfire distribution in the Idyllwild quandrangle, San Jacinto Mountain, California. Photogrammetric Engineering and Remote Sensing 56, 1507–1513. Cliff, A. D. and Ord, J. K. (1973). Spatial Autocorrelation. London: Pion. Craig, R. G. and Labovitz, M. L. (1980). Sources of variation in Landsat autocorrelation. Proceedings of the 14th International Symposium on Remote Sensing of Environment. pp. 1755–1767. Costa Rica: San Jose. Elkie, P. C., Rempel, R. S. and Carr, A. P. (1999). Patch Analyst User Manual: A Tool for Quantifying Landscape Structure. NWST Technical Manual TM002. Ontario. ESRI Environmental Services Research Incorporated. (1998). ARC/INFO v. 7.2.1. Redlands, CA: Environmental Services Research Incorporated. ESRI Environmental Services Research Incorporated (1999). ArcView v. 3.2. Redlands, CA: Environmental Services Research Incorporated. Forman, R. T. T. (1995). Land Mosaics: The Ecology of Landscapes and Regions. Cambridge: Cambridge University Press. Forman, R. T. T and Godron, M. (1981). Patches and structural components for a landscape ecology. Bioscience 31(10), 733–739. Fotheringham, A. S., Charlton, M. and Brunsdon, C. (1994). The geography of parameter space: an investigation of spatial non-stationarity. In Evaluating Spatial Modelling Techniques (P. A. Walker and R. J. Aspinall, eds), pp. 42–52. CSIRO report on a meeting held at the Macualay Land Use Research Institute, Aberdeen, Scotland, Sept 1994. Getis, A. and Ord. J. K. (1996). Local spatial statistics: an overview. In Spatial Analysis: Modelling in a GIS Environment (P. Longley and M. Batty, eds), pp. 261–277. Geoinformatics International. Goodchild, M. F. (1986). Spatial Autocorrelation. Catmog 47. Norwich: GeoBook. Haines-Young, R. and Chopping, M. (1996). Quantifying landscape structure: a review of landscape indices and their application to forested landscapes. Progress in Physical Geography 20(4), 418–445. Haynes, C. D. (1985). The pattern and ecology of munwag: traditional aboriginal fire regimes in north central Arnhem Land. Proceedings of the Ecological Society of Australia 13, 203–214.
Local measures of spatial autocorrelation Hobbs, R., Saunders, D. A. and Arnold, G. W. (1993). Integrated landscape ecology: a Western Australian perspective. Biological Conservation 64, 231–238. Innovative GIS Solutions Inc. (1999). FragstatsŁ Arc: Landscape Structure and Spatial Pattern Analysis for ArcInfo. URL http://blaze.innovativegis.com/products/ fragstatsarc/ Kienast, F. (1993). Analysis of historical landscape patterns with a geographical information system – a methodological outline. Landscape Ecology 8(2), 103–118. Labovitz, M. L., Toll, D. L. and Kennard, R. E. (1980). Preliminary Evidence for the Influence of Physiography and Scale upon the Autocorrelation Function of Remotely Sensed Data. NASA. TM 82064 Goodard Space Flight Center, Greenbelt MD. Legendre, P. and Fortin, M. J. (1989). Spatial pattern and ecological analysis. Vegetatio 80, 107–138. Ludwig, J. A., Eager, R. W., Williams, R. J. and Lowe, L. M. (1999). Declines in vegetation patches, plant diversity and grasshopper density near cattle watering-points in the Victoria River District, Northern Australia. The Rangeland Journal 21(1), 135–49. McGarigal, K. and Marks, B. J. (1994). FRAGSTATS: Spatial Pattern Analysis Program for Quantifying Landscape Pattern. Forest Science Dept, Oregon State University, Corvallis. Oregon. McIntyre, S. and Barrett, F. W. (1992). Habitat variegation, an alternative to fragmentation. Conservation Biology 6, 146–147. Musick, H. B. and Grover, H. D. (1991). Image textural measures as indices of landscape pattern. In Quantitative Methods in Landscape Ecology (M. G. Turner and R. H. Gardner, eds), pp. 77–103. New York: SpringerVerlag. Pearson, D. M. (1998). The Impact of Human Activity on Landscape Diversity in Space and Time. PhD thesis. Dept of Geography, ANU, Canberra, Australia. Pearson, D. M. and Riley, J. (2001). Spatial analysis of grassland patterning in the Yinberrie Hills region of the Northern Territory using field data and IKONOS imagery. Proceedings of the Northern Australian GIS and Remote Sensing Conference, July 3–5, Darwin 2001. Pickup, G., Bastin, G. N. and Chewings, V. H. (2000). Measuring rangeland vegetation with high-resolution airborne videography in the blue-near infrared spectral region. International Journal of Remote Sensing 21, 339–351. Price, O. and Bowman, D. M. J. S. (1994) Firestick forestry: a matrix model in support of skilful fire management of Callitris intratropica by north
95
Australian Aborigines. Journal of Biogeography 21, 573–580. Pyne, S. J. (1991). Burning Bush: A Fire History of Australia. New York: Henry Holt and Co. Qi, Y. and Wu, J. (1996). Effects of changing spatial resolution, on the results of landscape pattern analysis using spatial autocorrelation indices. Landscape Ecology 11(1), 39–49. Russell-Smith, J., Allan, G., Thackway, R., Rosling, T. and Smith, R. (2000). Fire management and savanna landscapes in Northern Australia. In Fire and Sustainable Agriculture and Forestry Development in Eastern Indonesia and Northern Australia (J. Russell-Smith, G. Hill, S. Djoeroemana and B. Myers, eds), pp. 95–101. Canberra: Australian Centre for International Agricultural Research. Stow, D. A. (1993). The role of geographic information systems for landscape ecological studies. In Landscape Ecology and GIS (R. Haines-Young, D. R. Green and S. H. Cousins, eds), pp. 11–21. London: Taylor and Francis. Tidemann, S. (1996). Causes of the decline of the Gouldian finch (Erthrura gouldia). Bird Conservation International 6, 49–61. Turner, M. G. (1990). Spatial and temporal analysis of landscape patterns. Landscape Ecology 4(1), 21–30. Turner, M. G., O’Neill, R. V., Gardner, R. H. and Milne, B. T. (1989). Effects of changing spatial scale on the analysis of landscape pattern. Landscape Ecology 3(3/4), 153–163. Walsh, S. J. and Davis, F. W. (1994). Applications of remote sensing and geographic information systems in vegetation science: introduction. Journal of Vegetation Science 5, 610–615. Whitehead, P. and Gorman, J. T. (1999). Defining and measuring the health of savanna landscapes: a north Australian perspective. In Proceedings from Ecosystem Health Workshop. 28–29 May 1999, Darwin. Williams, R. J., Cook, G. D., Ludwig, J. A. and Tongway, D. J. (1997). Torch, trees, teeth and tussocks: disturbance in the tropical savannas of the Northern Territory (Australia). In Frontiers in Ecology: Building the Links (N. Klump and I. Lunt, eds), pp. 55–66. Oxford: Elsevier Science. Woinarski, J. C. Z. and Whitehead, P. (2000). Conservation Management at a Landscape Scale in Northern Australia – Coping with Uncertainty and Change. Proceedings of the Australasian Wildlife Management Society Conference, Darwin, Dec 1999. Woodcock, C. E. and Strahler, A. H. (1987). The factor of scale in remote sensing. Remote Sensing of the Environment 21, 311–332.
The application of local measures of spatial autocorrelation for describing pattern in north Australian landscapes Diane M. Pearson School of Biological, Chemical and Environmental Science, Faculty of Science Information Technology and Education, Northern Territory University, Darwin, NT 0909, Australia Received 26 October 2000; accepted 2 October 2001
This paper tests the use of a spatial analysis technique, based on the calculation of local spatial autocorrelation, as a possible approach for modelling and quantifying structure in northern Australian savanna landscapes. Unlike many landscapes in the world, northern Australian savanna landscapes appear on the surface to be intact. They have not experienced the same large-scale land clearance and intensive land management as other landscapes across Australia. Despite this, natural resource managers are beginning to notice that processes are breaking down and declines in species are becoming more evident. With future declines of species looking more imminent it is particularly important that models are available that can help to assess landscape health, and quantify any structural change that takes place. GIS and landscape ecology provide a useful way of describing landscapes both spatially and temporally and have proved to be particularly useful for understanding vegetation structure or pattern in landscapes across the world. There are many measures that examine spatial structure in the landscape and most of these are now available in a GIS environment (e.g. FRAGSTATSŁ ARC, r.le, and Patch Analyst). All these methods depend on a landscape described in terms of patches, corridors and matrix. However, since landscapes in northern Australia appear to be relatively intact they tend to exist as surfaces of continuous variation rather than in clearly defined homogeneous units. As a result they cannot be easily described using entity-based models requiring patches and other essentially cartographic approaches. This means that more appropriate methods need to be developed and explored. The approach examined in this paper enables clustering and local pattern in the data to be identified and forms a generic method for conceptualising the landscape structure where patches are not obvious and where boundaries between landscape features are difficult to determine. Two sites are examined using this approach. They have been exposed to different degrees of disturbance by fire and grazing. The results show that savanna landscapes are very complex and that even where there is a high degree of disturbance the landscape is still relatively heterogeneous. This means that treating savanna landscapes as being made up of homogeneous units can limit analysis of pattern, as it can over simplify the structure present, and that methods such as the autocorrelation approach are useful tools for quantifying the variable nature of these landscapes. 2002 Academic Press
Keywords: Landscape structure, spatial analysis, landscape ecology, continuous variation, spatial autocorrelation, northern Australian savanna landscapes.
Introduction Although the tropical savanna landscapes of northern Australia appear to be intact they are under considerable pressure from anthropogenic activity. The result is subtle alterations to the landscape
Email: [email protected] 0301–4797/02/010085C11 $35.00/0
that are resulting in declines in certain species (Price and Bowman, 1994; Braithwaite and Griffiths, 1996) and progressive changes in structure that are thought to be detrimental (Woinarksi and Whitehead, 2000). One suite of species that are affected by these changes are granivorous birds, most notably the Gouldian finch (Erythrura gouldiae), that are diminishing in numbers across northern Australia (Tidemann, 1996). 2002 Academic Press
86
D. M. Pearson
The structural modifications that these landscapes have experienced are thought to be associated with pressures from grazing and changes in fire regimes. Fire has for a long time had a dominant influence on northern Australian savanna landscapes (Pyne, 1991) and many of the native species of vegetation are adapted to cope with fire at some stage in their life cycle and are often reliant upon it. Prior to European settlement burning was practised by Aboriginal people for cultural reasons, hunting purposes, and in order to provide good grazing habitat for the wildlife they hunted (Haynes, 1985; Braithwaite and Roberts, 1995). This created a mosaic of burnt and unburnt habitat in the landscape. However, since European settlement, fire regimes have changed, and now vast tracts of the landscape are burnt on an annual basis (Russell-Smith et al., 2000). It is currently thought that recent fire regimes in parts of northern Australia have been too frequent and too expansive (particularly when held in the late dry season) to ensure adequate regeneration of some species and to maintain biodiversity (RussellSmith et al., 2000). Added to this, in recent years northern Australian savanna landscapes have been put under greater pressure from increased cattle grazing intensities (Ludwig et al., 1999). In order to achieve sustainable use of these rangelands a better understanding of their ecology, particularly their spatial and temporal structure and their functional roles, is required (Ludwig et al., 1999). Since the spatial arrangement of landscape elements and the relationship between these elements, (the complexity between and within habitat patches), and how these change over time, is thought to influence the distribution and abundance of species, it is becoming increasingly important for natural resource managers in northern Australia to be able to successfully model landscape pattern and determine the ‘healthiness’ of a landscape. A healthy landscape has been defined by the Corporate Research Centre for Sustainable Tropical Savannas as one that can (i) maintain basic landscape functionality (e.g. nutrient cycling, water capture, food and shelter for fauna) at different spatial scales, (ii) maintain viable populations of native species at appropriate spatial and temperate scales, and (iii) reliably meet the needs of people over the long term (Whitehead and Gorman, 1999). Landscape ecology together with Geographic Information Systems (GIS) technology provides a framework within which one can examine spatial and temporal differences and changes in structure and function at the landscape scale, and relate these changes to modifications in the pattern of
land use. These methods and techniques have been used extensively across the world (Stow, 1993; Bridgewater, 1993; Kienast, 1993; Walsh and Davis, 1994; McGarigal and Marks, 1994). To understand structural changes in the landscape a number of measures of landscape pattern and diversity have been devised and a comprehensive review of these indices are reported in Haines-Young and Chopping (1996). The majority of these measures are based on patch theory, which relies on the identification of the three basic landscape features – patches, matrix and corridors (Forman, 1995) by some form of land cover classification. Many of these indices are now contained within software packages which are fully integrated with GIS such as FRAGSTATSŁ ARC (Innovative GIS Solutions Inc, 1999), r.le (Baker and Cai, 1992) and Patch Analyst (Elkie et al., 1999). Most of these measures take the form of indices, which means that they describe pattern through the calculation of a single number. Most of the commonly used indices have been developed for cultural landscapes in Europe and North America where agricultural development and management over centuries has produced a richly mosaiced landscape dominated by fields, field margins, woodland remnants and other discrete types of land cover. Since traditional approaches to landscape ecology have tended to be based on the identification of patch-matrix-corridor features (Forman and Godron, 1981), in cartographic terms, this means partitioning the landscape into homogeneous units with distinct boundaries, and the geometry (area and perimeter) of these homogeneous units are used as inputs to the computation of landscape metrics. This approach works well when applied to cultural landscapes in Europe and America, and even in parts of Australia, such as the wheatbelt of Western Australia (Hobbs et al., 1993), where landscape features can be clearly identified. However, it is less relevant for modelling areas where there is a gradation of land cover and where boundaries between landscape features are not clearly identified (McIntyre and Barrett, 1992). Such gradations of land cover are evident in the tropical savannas found in northern Australia. This is also a problem encountered in other savanna landscapes in the world and in other landscapes were structural modifications are subtle, like in forested landscapes were selective logging has taken place rather than large-scale clearance (Pearson, 1998). This paper investigates the use of an alternative approach to quantifying structure and structural changes based on modelling pattern
Local measures of spatial autocorrelation
through the use of a measure of local spatial autocorrelation, integrated within a GIS (ARC/INFO (ESRI, 1998)) (Pearson, 1998). For comparison, the results of modelling landscape pattern using local autocorrelation are compared with a more traditional way of representing the pattern present in the landscape (a simple classification of land cover) that was produced by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) (Eager, pers. comm.; Pickup et al., 2000).
Study area The study sites examined in this paper are taken from the Kidman Springs Research Station which is situated in the Victoria River District (VRD) of the Northern Territory and lies about 700 km south of Darwin (see Figure 1). The Station is managed as a research pastoral property owned by the Department of Primary Industry and Fisheries and consists of rangelands/tropical savannas that are typical of the region. They are under considerable pressure from grazing and management for grazing using fire.
In the Kidman Springs area the median rainfall is approximately 640 mm, which is largely, summer (wet season) dominant (Ludwig et al., 1999). Trees and perennial grasses in the tropical north of Australia rely on being able to capture and store water and nutrients during the wet season for their growth and survival in the dry season. If plants are unable to capture and store water and nutrients they will become stressed, especially in the late dry season, which will result in increased plant mortality, particularly amongst those plants which are susceptible to grazing and fire (Williams et al., 1997; Ludwig et al., 1999). The pattern or structure of vegetation patches plays an important role in capturing and conserving important resources. Where landscapes have lost significant vegetation patches (particularly perennials) they have a reduced capacity to capture and store scarce resources. This affects the overall condition or healthiness of the landscape, which in turn influences its capacity to sustain viable populations of flora and fauna (Ludwig et al., 1999). The loss of significant vegetation patches can also have a knock on effect in terms of a loss of diversity of species of fauna that depend on these plants for their survival, and where overgrazing takes place
Darwin
Kidman Springs
Figure 1.
87
Location of Kidman Springs study site in the Victoria River District of the Northern Territory, Australia.
88
D. M. Pearson
in the rangelands, soil erosion can result, which in turn can affect the stability of the landscape and its sustainability for supporting biodiversity. Cattle tend to focus grazing and trampling near to watering points, as a result the distance from a watering point is used in this study as a surrogate for a gradient in grazing pressure (Ludwig et al., 1999). Sites were chosen so that differences in pattern could be observed relative to exposure to fire and heavy grazing.
Data The data used in this case study consist of two images of the Kidman Springs area captured using airborne videography, which have a resolution of 0Ð2 m. One of the sites depicted by an image is very close to a watering hole (KS1) the other is in an exclosure (KS5) (grazing and fire have been excluded from this area for about 25 years). The dominant vegetation in both sites is low, open eucalypt savanna with the typical grass species being perennial black spear grass (Heteropogon contortus), curly bluegrass (Dichanthium fecundum), biennial limestone grass (Enneapogon polyphyllus) and the annual, false couch (Brachyachne convergens) (Ludwig et al., 1999). Both sites are on the same land unit type. This is described as ‘gently undulating plains with calcareous red earth soils’ (Ludwig et al., 1999). Videography imagery was captured individually for the two sites for four bands (blue, green, red and infrared). A simple sequential classification procedure was employed by CSIRO using transforms and indices rather than raw bands to sequentially identify the five main land cover types in the area i.e. shadow, tree/shrub canopy, bare soil, perennial patches, and annuals/litter (Eager pers. comm.; Pickup et al., 2000). This classification is included as ancillary data in this study for comparative purposes to contrast more traditional ways of representing landscape pattern with the alternative approach put forward in this paper.
Methodology The approach examined in this study is to apply a technique for modelling pattern that was thought to be better suited for understanding the spatial heterogeneity or pattern in landscapes where cartographic description fails to capture the variability present (Pearson, 1998). This approach is
consistent with that suggested by Musick and Grover (1991) who recommend applying textural measures, which take account of spatial variation via analysis of brightness values in the data as a method of modelling spatial pattern present in landscapes. The approach explored is geostatistical and applies a local measure of spatial autocorrelation based on the statistic Geary’s C. This type of mathematical measure is referred to as a structure function (Legendre and Fortin, 1989). Measures of spatial autocorrelation (Cliff and Ord, 1973) work by examining how objects at one location are similar to objects located nearby (Goodchild, 1986), that is, they look for spatial dependence in the data (similarity as a function of distance) (Chou et al., 1990). If features situated close together have similar attribute information, then the pattern in the data can be described as exhibiting positive autocorrelation. When features close together are more dissimilar in attribute value than features further away, pattern in the data is negatively autocorrelated. Zero autocorrelation exists when attributes or their values are independent of location (Goodchild, 1986). Fieldwork shows that boundaries between patches are difficult to distinguish in this landscape with large areas of land existing as a mixture of annuals and perennials grading into one another (this is particularly the case in KS1). The grading between vegetation classes makes it difficult to clearly define boundaries around patches in these sites. This means that classifying the landscape into broad groups can over-simplify the ground conditions. As a result, in order to prevent simplifying a complex landscape the raw spectral data were used for analysis with the autocorrelation approach, thus no objects such as land cover polygons are pre-defined in the data prior to analysis. The approach has previously been applied to research conducted on New South Wales (NSW) forested landscapes (Pearson, 1998). Although the forested landscapes in NSW are different to the savanna landscapes found in northern Australia, it is thought that both types of landscape are less inclined to exhibit the clearly defined patch, matrix and corridor features usually examined in landscape ecological analysis. In contrast, they demonstrate gradations between vegetation classes across the landscape. Conventional measures of spatial autocorrelation in a GIS operate on a global basis, thus they perform simultaneous measurements of spatial dependence for all locations within the dataset and return a single value for that data set. Getis
Local measures of spatial autocorrelation
and Ord (1996) recommend the use of local statistics, which work by measuring spatial association between one pixel and its neighbours up to a certain sampling distance. An ARC/INFO AML (ESRI, 1998) program was written to produce a local measure of Geary’s C (Pearson, 1998). It operates on the principle of examining the relationship between a variable measured at a given point and the same variable measured at points some distance from the original point. This approach is similar to a method used by Fotheringham et al. (1996) for investigating spatial non-stationarity. This methodology applies the principles of landscape ecology to cells rather than patches. It enables one to look at variation not only in space and time, but also to examine the effect of changing the scale of analysis on the output, by varying the grain and extent (Turner, 1990). Since four bands of spectral information make up each videography image and the autocorrelation routine can only run on one surface at a time, a new GRID surface was created on which to perform the analysis. This was made up of the average spectral reflectance values for the four bands of data. The effect of changing the scale of analysis has an important impact on the results because all measures of landscape pattern are scale dependant (Haines-Young and Chopping, 1996), and results are meaningless unless made with reference to the scale of measurement (Turner et al., 1989). Qi and Wu (1996) looked at the effects of changing spatial resolution on the results of landscape pattern analysis using spatial autocorrelation indices. They showed that the degree of autocorrelation present in pattern is dependant on scale and that changing the scale effects the amount of autocorrelation found in pattern. Appropriate scales of analysis tend to be a function of the type of environment being examined and the sort of information that is required (Woodcock and Strahler, 1987). Craig and Labovitz (1980) and Labovitz et al. (1980) measured autocorrelation in Landsat MSS and TM images and found it to be higher in images with a fine spatial resolution than in images of much coarser resolution. Performing analysis on grids of varying size, allows the effect of changing the spatial resolution on the result to be assessed. The importance of scale is addressed in this autocorrelation measure by altering the sample size (i.e. the distance over which spatial dependence is calculated or the area that the moving window covers), and also the cell size (i.e. the resolution of analysis – this is changed using the GRID SETCELL function in ARC/INFO (ESRI, 1998)). These steps were built into the ARC/INFO
89
AML (ESRI, 1998) program so the user could run the analysis at a range of cell and window sizes and compare the results. The results are new sets of surfaces for each of the scales and sample sizes examined. Changing the scale of analysis helps to determine appropriate scales at which certain landscape features operate. In order to evaluate the impact of changing scale of analysis on the results the analysis was conducted at cell sizes starting from 0Ð2 m to 1 m and windows of analysis ranging from 1 m to 100 m. Determining the most appropriate scale of analysis is based on the user’s interpretation of the results in conjunction with patch location and size information collected in the field. In order to get some idea of the accuracy of the autocorrelation approach in terms of its ability to determine patches in the landscape, ArcView (ESRI, 1999) GIS was used to overlay the locations of patches recorded from fieldwork on top of the resulting autocorrelation surfaces. These figures were compared with overlaying the same field data on the results of the CSIRO classification for the two study sites.
Results Conducting the analysis at a range of cell sizes and windows of analysis, then examining the output overlaid with the field data in ArcView (ESRI, 1999) GIS, enabled appropriate scales of analysis to be determined for each of the study sites. The results showed that a cell size of as fine a resolution as possible was most appropriate for both sites, as this was important for maintaining detail in the image and enabling small scale pattern in the landscape to be determined. The finest resolution possible with the videography imagery used in this study is the pixel size, which is 0Ð2 m. Figure 2 shows the effect of changing the window of analysis on the output from KS1. Similar patterns were evident for KS5. Small windows of analysis produced a surface that was made up of a pattern that was very speckled in appearance and bore no resemblance to the pattern actually present in these landscapes. As the window size increased a pattern that was more consistent with vegetation patches recorded in the field emerged. Up to a certain scale, areas of positive autocorrelation start to correspond to patches recorded in the field, which could be either patches of perennials, annuals or tree canopies. However, beyond a certain window size the data structure becomes overly simplified and the influence of the edge
90
D. M. Pearson
Figure 2. The effect of changing the window of analysis on the results of autocorrelation on KSI. Window sizes (in m): (a) 4ð4; (b) 16ð16; (c) 32ð32; (d) 60ð60.
of the data becomes more evident until a point where the amount of useable data is reduced to a very narrow strip. Analysis at different scales revealed that for site KS1 a window size of 60 m ð 60 m was appropriate for modelling landscape pattern, whilst a window size of 48 m ð 48 m was deemed to be appropriate for modelling of pattern in KS5. Figures 3 and 4 show that, at the scales listed above, the autocorrelation approach produces surfaces that are able to identify some of the main patches detected by fieldwork. Patches of similar land cover type are represented as areas that are positively autocorrelated, while the surrounding areas that are more variable, are represented by regions of negative autocorrelation. What occurs is that positive autocorrelation corresponds to patches of perennials or annuals identified from fieldwork while negative autocorrelation tends to correspond to where there is a mixture of grass species (annuals and perennials grading into one another) or where there is a mix of bare ground and some grassy tussocks.
The pattern depicted in the autocorrelation surfaces seems to resemble the pattern recorded in the field more closely than the results produced from applying a simple classification to the data (see Figures 5 and 6). A comparison with the CSIRO classification data at KS1 and KS5 shows that the autocorrelation surfaces give more indication of the variability present in the landscape and therefore do not over simplify what is in effect a very complex landscape by treating the area as being capable of being defined as a series of homogeneous units. Table 1 shows that for KS1 the percentage of the area that is positively autocorrelated is 18Ð7% while 77Ð7% of the area is negatively autocorrelated. This compares with 22Ð7% and 73Ð3% respectively for KS5. These figures provide further evidence of the amount of heterogeneity present in both these landscapes. The percentage of the area that is negatively autocorrelated in KS1 is particularly a surprise if one views the classified image for this landscape. The CSIRO classification for KS1 depicts that the landscape is made up of quite large homogeneous areas; particularly the area
Local measures of spatial autocorrelation
91
Figure 3. Spatial autocorrelation surface for KSI (0Ð2 m cell size and 60ð60 m window) with location of patches recorded in the field.
Figure 4. Spatial autocorrelation surface for KS5 (0Ð2 m cell size and 48ð48 m window) with location of patches recorded in the field.
described as bare ground. The results of applying the autocorrelation approach therefore reveals that the Kidman Springs landscapes are much more variable spatially than a simple classification will allow for.
Discussion The results show that the autocorrelation approach is capable of describing more of the spatial
variability present at each of the sites examined in this paper than is evident by performing a simple classification on the same sites. This result is consistent with the idea that grassy landscapes tend to exhibit continuous variation, which make them difficult to represent in a two-dimensional surface made up of homogeneous units with distinct boundaries (Pearson and Riley, 2001). Therefore, to define patchiness or pattern in the landscape using the autocorrelation approach appears to be one way of
92
D. M. Pearson
Figure 5.
Classification of KS1 and the location of patches recorded in the field.
Figure 6.
Classification of KS5 and the location of patches recorded in the field.
addressing the problem of high spatial variability in grassy landscapes and producing surfaces that give a more realistic representation of the pattern present. The advantages of modelling pattern using spatial autocorrelation over other ways of quantifying landscape structure are that it allows environmental variation to be represented as a set of distinct descriptions at grid cells across the image and reduces the problems associated with indeterminate boundaries between landscape features. Other advantages that exist by defining landscape
Table 1. Percentage of total area for each site that is represented by each level in the autocorrelation surfaces Site
Percent positively autocorrelated
Percent zero autocorrelated
Percent negatively autocorrelated
KS1 KS5
18Ð7 22Ð7
3Ð6 3Ð6
77Ð7 73Ð7
pattern by the autocorrelation approach rather than classifying landscapes to produce surfaces made up of homogeneous regions, are that many
Local measures of spatial autocorrelation
of the fine scale patches or pattern that may be important to grassy landscapes, particularly for species like granivorous birds, may not be detected if the landscape becomes too simplified by a broad classification. For example, what is described as homogeneous ‘bare ground’ in the CSIRO model of pattern shows up to be quite spatially variable using the autocorrelation approach. Closer inspection of the field data shows that what is classified as bare ground is actually a mixture of tussocks of grass or small, close to the ground patches of grass species interspersed with bare ground. Therefore, the spectral response from remote sensing is expected to be mixed in this region so it is not appropriate to treat this area as being homogeneous if one needs to understand fine scale structure. Fieldwork also showed that KS1 has more intermixing of annuals and perennials than KS5, thus making boundaries between patches difficult to determine in the field and producing a heterogeneous effect at the fine scale, which explains the large percentage of the area that was negatively autocorrelated in the KS1 autocorrelation surface. Field investigation of the two sites showed that in comparison to KS1, KS5 had more dense patches and tended to have larger patches of mostly perennials with only small patches of annuals present. This again explains why KS5, which although appears on the surface to be more heterogeneous than KS1, actually produced autocorrelation results that show a smaller percentage of the area being negatively autocorrelated than KS1. Differences in structure observed both by fieldwork and from the autocorrelation surfaces are probably related to higher disturbance factors in KS1, in particular those caused by intensive cattle grazing and trampling close to the watering hole. KS5 has been relatively undisturbed by either fire or grazing. The fact that KS1 is seen to be more heterogeneous at fine resolution as a result of disturbance could indicate that this part of the landscape is actually more unstable than KS5, making the approach a possible indicator of landscape condition/health. The spatial autocorrelation approach is also particularly useful for showing the effect of scale on landscape studies and how different land use features operate at different scales and therefore require different scales of analysis, (KS1 required a larger window of analysis to identify the large significant patches than KS5). Patches identified by fieldwork are only found with this method by applying the analysis at a specific spatial scale, resolution or window size. The results indicate how spatial scale plays an integral role in determining the results from analysis of pattern or structure
93
present in the landscape. However, it could also be seen as being a limitation with the autocorrelation approach as, in order to ensure meaningful results, the resolution of analysis needs to be carefully selected. Another limitation of this approach is that although viewing the autocorrelation surfaces in the GIS shows that they succeeded in identifying many of the significant patches that were detected from fieldwork, there are still some discrepancies between patch locations in the field and those on the surfaces. More fieldwork is needed to be able to explain these anomalies fully and to understand why some patches are detected and others are missed by this approach. Despite some of the limitations, the autocorrelation approach has the potential to be a valuable tool for use in landscape scale studies, as it may help the user to select appropriate scales of analysis for particular landscape types. Issues of scale of analysis and detecting scales of operation of landscape features are major problems for researchers interested in landscape pattern that have not been adequately addressed up to now.
Conclusion Change to landscape structures that adversely affects their functioning can have serious implications for the health of a landscape and its biodiversity. More survey work is needed to fully understand the association between patterns of vegetation and landscape health. GIS and remote sensing technologies, coupled to modelling procedures such as the one described in this paper, can form useful surrogates for monitoring the status and condition of landscapes and therefore have the potential to be useful indicators of landscape health (Aspinall and Pearson, 2000) in the tropical savannas. The technique described in this paper is currently being applied to try to understand grassland patterns and how they relate to the sustainable management of granivorous birds in northern Australia (Pearson and Riley, 2001). Such studies involve many complexities, but the ability to successfully model some of the complex spatial associations and patterns in landscapes will ultimately help us to understand the impacts of human activity on northern Australian landscapes, and to explain recent declines of species such as the Gouldian finch from savanna landscapes. The application of geostatistical techniques such as the autocorrelation approach examined
94
D. M. Pearson
in this paper can be seen as being a positive step towards successfully modelling heterogeneous savanna landscapes. It forms a generic method of conceptualising the landscape structure with a particular scale, in areas where patch structure is not obvious and where boundaries between landscape features are difficult to determine, thus helping to deal with problems associated with indeterminate boundaries (Burrough, 1996) that are a major issue in places like northern Australia. Classification of remotely sensed imagery for grassland species is very difficult because of the complex and heterogeneous nature of the landscape. Dividing the landscape into broad categories, as usually happens in classification, enables some pattern to be identified, but the result may end up excluding pattern at a smaller scale which could in fact be the structure that is most important for species that inhabit these landscapes, such as granivorous birds (Pearson and Riley, 2001). Since the autocorrelation approach does not impose structure prior to analysis, none of the variability present in the data is simplified before analysis. As a result it may have an advantage over the currently available procedures for quantifying landscape structure, which fail to describe north Australian savanna landscapes adequately because of their highly variable nature (Pearson and Riley, 2001).
Acknowledgments I would like to thank the Northern Territory University for funding the project. CSIRO and the Corporate Research Centre for the Sustainable Development of Tropical Savannas (TS CRC) especially John Ludwig and Robert Eager, provided data and encouragement for this project. Thanks also to everyone at the Kidman Springs Research Station for their hospitality, two anonymous referees. Julian Gorman for assistance with the fieldwork and discussions on the research, and Richard Aspinall for his comments on earlier drafts of this paper.
References Aspinall, R. J. and Pearson, D. M. (2000). Integrated geographical assessment of environmental condition in water catchments: linking landscape ecology, environmental modeling and GIS. Journal of Environmental Management 59(4), 299–319. Baker, W. L. and Cai, Y. (1992). The role programs for multiscale analysis of landscape structure using the GRASS geographical information system. Landscape Ecology 7(4), 291–302.
Braithwaite, R. W. and Griffiths, A. D. (1996). The paradox of rattus tunneyi: endangerment of a native pest. Wildlife Research 23, 1–21. Braithwaite, R. W. and Roberts, S. (1995). Aboriginal burning and conservation management in Kakadu National Park, tropical Australia. Wildfire September, 29–38. Bridgewater, P. B. (1993). Landscape ecology, geographic information systems and nature conservation. In Landscape Ecology and GIS (R. Haines-Young, D. R. Green and S. H. Cousins, eds), pp. 23–36. London: Taylor and Francis Ltd. Burrough, P. A. (1996). Natural objects with indeterminate boundaries. In Geographic Objects with Indeterminate Boundaries (P. A. Burrough and A. U. Frank, eds), pp. 3–28. London: Taylor and Francis Ltd. Chou, Y. A., Minnich, R. A., Salatar, L. A., Power, J. D. and Dezzani, R. J. (1990). Spatial autocorrelation of wildfire distribution in the Idyllwild quandrangle, San Jacinto Mountain, California. Photogrammetric Engineering and Remote Sensing 56, 1507–1513. Cliff, A. D. and Ord, J. K. (1973). Spatial Autocorrelation. London: Pion. Craig, R. G. and Labovitz, M. L. (1980). Sources of variation in Landsat autocorrelation. Proceedings of the 14th International Symposium on Remote Sensing of Environment. pp. 1755–1767. Costa Rica: San Jose. Elkie, P. C., Rempel, R. S. and Carr, A. P. (1999). Patch Analyst User Manual: A Tool for Quantifying Landscape Structure. NWST Technical Manual TM002. Ontario. ESRI Environmental Services Research Incorporated. (1998). ARC/INFO v. 7.2.1. Redlands, CA: Environmental Services Research Incorporated. ESRI Environmental Services Research Incorporated (1999). ArcView v. 3.2. Redlands, CA: Environmental Services Research Incorporated. Forman, R. T. T. (1995). Land Mosaics: The Ecology of Landscapes and Regions. Cambridge: Cambridge University Press. Forman, R. T. T and Godron, M. (1981). Patches and structural components for a landscape ecology. Bioscience 31(10), 733–739. Fotheringham, A. S., Charlton, M. and Brunsdon, C. (1994). The geography of parameter space: an investigation of spatial non-stationarity. In Evaluating Spatial Modelling Techniques (P. A. Walker and R. J. Aspinall, eds), pp. 42–52. CSIRO report on a meeting held at the Macualay Land Use Research Institute, Aberdeen, Scotland, Sept 1994. Getis, A. and Ord. J. K. (1996). Local spatial statistics: an overview. In Spatial Analysis: Modelling in a GIS Environment (P. Longley and M. Batty, eds), pp. 261–277. Geoinformatics International. Goodchild, M. F. (1986). Spatial Autocorrelation. Catmog 47. Norwich: GeoBook. Haines-Young, R. and Chopping, M. (1996). Quantifying landscape structure: a review of landscape indices and their application to forested landscapes. Progress in Physical Geography 20(4), 418–445. Haynes, C. D. (1985). The pattern and ecology of munwag: traditional aboriginal fire regimes in north central Arnhem Land. Proceedings of the Ecological Society of Australia 13, 203–214.
Local measures of spatial autocorrelation Hobbs, R., Saunders, D. A. and Arnold, G. W. (1993). Integrated landscape ecology: a Western Australian perspective. Biological Conservation 64, 231–238. Innovative GIS Solutions Inc. (1999). FragstatsŁ Arc: Landscape Structure and Spatial Pattern Analysis for ArcInfo. URL http://blaze.innovativegis.com/products/ fragstatsarc/ Kienast, F. (1993). Analysis of historical landscape patterns with a geographical information system – a methodological outline. Landscape Ecology 8(2), 103–118. Labovitz, M. L., Toll, D. L. and Kennard, R. E. (1980). Preliminary Evidence for the Influence of Physiography and Scale upon the Autocorrelation Function of Remotely Sensed Data. NASA. TM 82064 Goodard Space Flight Center, Greenbelt MD. Legendre, P. and Fortin, M. J. (1989). Spatial pattern and ecological analysis. Vegetatio 80, 107–138. Ludwig, J. A., Eager, R. W., Williams, R. J. and Lowe, L. M. (1999). Declines in vegetation patches, plant diversity and grasshopper density near cattle watering-points in the Victoria River District, Northern Australia. The Rangeland Journal 21(1), 135–49. McGarigal, K. and Marks, B. J. (1994). FRAGSTATS: Spatial Pattern Analysis Program for Quantifying Landscape Pattern. Forest Science Dept, Oregon State University, Corvallis. Oregon. McIntyre, S. and Barrett, F. W. (1992). Habitat variegation, an alternative to fragmentation. Conservation Biology 6, 146–147. Musick, H. B. and Grover, H. D. (1991). Image textural measures as indices of landscape pattern. In Quantitative Methods in Landscape Ecology (M. G. Turner and R. H. Gardner, eds), pp. 77–103. New York: SpringerVerlag. Pearson, D. M. (1998). The Impact of Human Activity on Landscape Diversity in Space and Time. PhD thesis. Dept of Geography, ANU, Canberra, Australia. Pearson, D. M. and Riley, J. (2001). Spatial analysis of grassland patterning in the Yinberrie Hills region of the Northern Territory using field data and IKONOS imagery. Proceedings of the Northern Australian GIS and Remote Sensing Conference, July 3–5, Darwin 2001. Pickup, G., Bastin, G. N. and Chewings, V. H. (2000). Measuring rangeland vegetation with high-resolution airborne videography in the blue-near infrared spectral region. International Journal of Remote Sensing 21, 339–351. Price, O. and Bowman, D. M. J. S. (1994) Firestick forestry: a matrix model in support of skilful fire management of Callitris intratropica by north
95
Australian Aborigines. Journal of Biogeography 21, 573–580. Pyne, S. J. (1991). Burning Bush: A Fire History of Australia. New York: Henry Holt and Co. Qi, Y. and Wu, J. (1996). Effects of changing spatial resolution, on the results of landscape pattern analysis using spatial autocorrelation indices. Landscape Ecology 11(1), 39–49. Russell-Smith, J., Allan, G., Thackway, R., Rosling, T. and Smith, R. (2000). Fire management and savanna landscapes in Northern Australia. In Fire and Sustainable Agriculture and Forestry Development in Eastern Indonesia and Northern Australia (J. Russell-Smith, G. Hill, S. Djoeroemana and B. Myers, eds), pp. 95–101. Canberra: Australian Centre for International Agricultural Research. Stow, D. A. (1993). The role of geographic information systems for landscape ecological studies. In Landscape Ecology and GIS (R. Haines-Young, D. R. Green and S. H. Cousins, eds), pp. 11–21. London: Taylor and Francis. Tidemann, S. (1996). Causes of the decline of the Gouldian finch (Erthrura gouldia). Bird Conservation International 6, 49–61. Turner, M. G. (1990). Spatial and temporal analysis of landscape patterns. Landscape Ecology 4(1), 21–30. Turner, M. G., O’Neill, R. V., Gardner, R. H. and Milne, B. T. (1989). Effects of changing spatial scale on the analysis of landscape pattern. Landscape Ecology 3(3/4), 153–163. Walsh, S. J. and Davis, F. W. (1994). Applications of remote sensing and geographic information systems in vegetation science: introduction. Journal of Vegetation Science 5, 610–615. Whitehead, P. and Gorman, J. T. (1999). Defining and measuring the health of savanna landscapes: a north Australian perspective. In Proceedings from Ecosystem Health Workshop. 28–29 May 1999, Darwin. Williams, R. J., Cook, G. D., Ludwig, J. A. and Tongway, D. J. (1997). Torch, trees, teeth and tussocks: disturbance in the tropical savannas of the Northern Territory (Australia). In Frontiers in Ecology: Building the Links (N. Klump and I. Lunt, eds), pp. 55–66. Oxford: Elsevier Science. Woinarski, J. C. Z. and Whitehead, P. (2000). Conservation Management at a Landscape Scale in Northern Australia – Coping with Uncertainty and Change. Proceedings of the Australasian Wildlife Management Society Conference, Darwin, Dec 1999. Woodcock, C. E. and Strahler, A. H. (1987). The factor of scale in remote sensing. Remote Sensing of the Environment 21, 311–332.