The evaluation of forest cover configuration depends on the availability of spatially referenced or mapped data. Most commonly, such data are derived from remote sensing, using either satellite-mounted or airborne sensors or aerial photography. For continental and global scale evaluation, satellite imagery is the most appropriate data source.
Satellite-mounted sensors measure radiation reflected from the Earth’s surface in a variety of spectral bands; different land-cover types have different characteristic reflectances and spectral signatures. Satellite sensors vary in the frequency with which they return to a given portion of the earth’s surface (temporal resolution), in the spatial resolution, or pixel size, of the data they provide, and in the spectral resolution or numbers and types of spectral bands in which data are recorded.
Satellite data are processed by adjusting them radiometrically to compensate for variations in atmospheric conditions, and then classifying the digital data by one of several methods. Classification can be achieved by plotting reflectance in particular spectral bands or band ratios such as NDVI (Normalised Difference Vegetation Index) and visually interpreting the results. Alternatively, satellite data can be classified digitally by grouping pixels with similar spectral characteristics, and by comparing their spectral responses to those of pixels from areas of known land cover (supervised classification). Additional power can be brought to the process by incorporating information on variation in spectral response of an area through time (seasonal and other changes) and ancillary data (on land forms, land use etc.). Expressing pixel composition in relation to mixtures of the spectral responses of possible component land covers (spectral mixture modelling) can also improve the resolution and accuracy of classification.
The processing of satellite data is expensive and time consuming, requiring sophisticated hardware and software to deal with the large volumes of digital data involved. Although very sophisticated and spatially detailed vegetation maps have been generated from satellite data, these have been confined to small areas because of the additional volume of data required for high spatial and spectral resolution processing.
The only currently available global land-cover data set derived from satellite data that have been processed in a consistent manner is the GLCCD, produced by the EROS Data Center and IGBP, from monthly averages of data from the AVHRR satellite during 1992-93 (Belward et al.,. 1999). These data have a spatial resolution of 1 km, and their relatively low spectral resolution has been compensated by the large amount of data on temporal variation that is available and by incorporating large amounts of ancillary data into the classification process (Loveland et al.,. 1999). At present, this is the only data set that could be used to evaluate forest fragmentation and regional and global scales.
Because of the coarse spatial resolution of the GLCCD, the data represent relatively coarse spatial mosaics and provide insufficient detail for certain kinds of analyses of forest configuration. There are also systematic errors in that result from the coarse spatial resolution: small non-forest patches in areas of high forest cover remain undetected as do small forest patches in landscapes with low forest cover. These phenomena lead respectively to over- and under- estimates of forest cover. Although these errors can be reduced by calibration against high-resolution data (Mayaux & Lambin 1995, 1997), it is not clear whether such a process has been applied to the GLCCD. Details of boundary configuration and interspersion of forest and non-forest land cover are lost at the lower spatial resolution. However, some consistent patterns that have ecological meaning for forest biodiversity do emerge. The following discussion focuses on the options for evaluating forest configuration that are appropriate for use with data like the GLCCD, but also includes examples of metrics that can be used productively with higher resolution data. The sensitivity of the different measures to data scale and resolution is discussed.
Defining ecosystems is an issue in the extent to which the analysis attempts to focus on individual forest types and their configuration within the landscape. Ideally one would look at the fragmentation parameters of each forest type within a mixed matrix of forest and non-forest separately, but in fact it is unlikely that the source data on forest type distribution can support this. It will probably be necessary to look just at forest cover in relation to the non-forest matrix, or perhaps at the fragmentation properties of rather broad or regional forest types with a minimum of overlap.
Scale is of course an issue. Fragmentation is differently determined for different components of the biota of a forest ecosystem. A path or deforested strip of a few metres width may be a significant barrier to an invertebrate, whereas a deforested strip of several km presents little obstacle to a forest-dwelling bird of prey. The data available for doing a globally consistent evaluation of forest fragmentation are 1 km resolution satellite data. Therefore, any fragmentation metrics that are derived from these data will represent the distribution of forests only at this coarse scale, and indeed might be better said to represent the configuration of forested areas in the landscape than of individual patches of forest. The effects of coarse resolution data on forest area estimates discussed above also apply to the estimation of forest fragmentation. Because of the problems of detecting small patches within very high or very low forest cover landscapes, fragmentation will be underestimated in areas of both very high and very low forest cover.
The question of what is natural or a baseline condition is also an issue in evaluating fragmentation of forests. Some forest types and regions are naturally more continuous than others. For example, forest close to latitudinal or other limits of its distribution has a tendency to be naturally patchy, as in the taiga/tundra transition of the boreal regions and some forest-savanna boundaries. Similarly forests that exist on mountains in non-forested landscapes because of the greater humidity at higher altitudes are inherently restricted in area and isolated from each other. The fragmented nature of these systems has a great significance for biodiversity that is entirely distinct from that of the fragmentation of continuous forest cover by human activity.
Some level of forest fragmentation may also result from natural disturbance and dynamic processes within forests. However, the scales of these processes are usually such that they will be beneath the resolution of any global or regional scale analyses. Although some kinds of storm damage or wildfire impact may be at scales that could be detected by these analyses, the issue of whether changes in forest fragmentation parameters are within or outside the range of natural variation is unlikely to arise.
A number of packages for use with geographic information systems (GIS) permit the analysis and characterisation of landscapes in terms of their patch composition, spatial relations and dynamics. One such package, FRAGSTATS (McGarigal and Marks 1995) is widely used for the description and analysis of landscape configuration. It offers a wide range of measures of varying complexity.
The choice of measures of forest fragmentation for use as indicators of forest capacity to retain biodiversity is dictated both by the source data and by the range of biological effects being targeted. Summary statistics of landscape metrics are of little use for predicting responses of individual species without more detailed information about both species requirements and environmental variation on the ground. However, to provide both an overview of forest status in relation to biodiversity and baselines to track changes that may affect forest biodiversity simple statistical expressions of forest configuration can be useful. It is also important that the metrics chosen be easily communicated and understood by the anticipated audience for the overview and monitoring, so conceptually complex indices are generally less useful.
Landscape ecological theory and GIS technology have generated a number of measures of the spatial distribution of habitat that express different aspects of its fragmentation in ways that relate to ecological processes. These have mostly been used in the evaluation of habitat and landscape processes at management unit scales. For example, Kramer (1997) used these analyses to describe landscape change as a result of management of two national parks in Costa Rica. She found that as forest cover increased, patch size increased and patch shapes became more compact for the shrinking pasture areas, but remained similar or became more complex over time for the expanding forest. Landscape diversity declined as the pasture areas shrank. In another example, Helmer (2000) used similar analyses to show that in the mountains of Costa Rica secondary forest is strongly associated with primary forest and occurs in smaller patches with more complex shapes. Logsdon et al.,. (2000) used FRAGSTATS to analyse and describe landscape composition, configuration and heterogeneity from remotely sensed data for a small part of central Amazonia. In one of the few studies that directly addresses the relationship between landscape parameters and potential management actions, Ranta et al.,. (1998) used landscape analysis to characterise the fragmentation of Atlantic rain forest in Brazil and to simulate the likely impacts of land use change or forest restoration. Still other studies have characterised forest fragmentation at broader scales. Skole & Tucker (1993) evaluated forest fragementation and potential edge impacts for the whole of the Amazon Basin, and substantial research has been done to characterise and describe patterns of fragmentation throughout the tropics (Jeanjean et al.,. 1994, 1995).
As these examples show, spatial analysis of landscapes and the metrics it produces have been the subject of much research and have been used widely for descriptive purposes. However, they represent a complex suite of ecological processes and effects that are beyond the detailed understanding of many decision-makers. Also, they often focus at scales (such as the 100 m edge effect postulated by Helmer 2000) that are beyond the resolution of currently available classified satellite data on landcover at regional and global scales. Furthermore, they have rarely been used to identify the likely relative value of different patches of habitat in the context of fragmentation, as a basis for monitoring changes in that condition or for making policy and management decisions.
Thus, there is a need for an easily understood summary index that can be used both as a basis for visualising the relative biodiversity preservation capacity of different forest areas and to establish a baseline for tracking forest landscape change. Both of these uses support decision-making and evaluation of policy effectiveness. Such an index needs to reflect all three types of fragmentation effects outlined above:
• area or patch size;
• interface with non-forest, or edge effects
• isolation from, or interconnection with other patches.
In this section of the paper, we develop a set of tools for quantifying forest fragmentation in relation to these three types of effects and displaying them in ways that can be meaningful to decision-makers. The use of each tool is illustrated using the forest cover of Paraguay (Figure 1) as an example data set. The data are part of a global data set derived from 1 km resolution AVHRR satellite data, and are therefore characteristic of forest cover data that could be used in global assessment and monitoring of forest fragmentation. Throughout the paper it will be useful to compare the mapped indices of fragmentation with this base map of forest cover to evaluate the additional information they convey and their potential utility for supporting decision making.
Area effects are most easily represented in terms of patch size. A GIS can be used to identify all patches of forest within the area of study, to measure their areas and assign them to a patch size class. If the data used are coarse resolution satellite data, this evaluation will identify contiguous blocks where forest is the predominant land cover, even if there may be some breaks in the forest cover at sub-pixel level.
The appropriate size class intervals can be selected empirically to provide a distribution with an easily analysed shape. The thresholds can also be adjusted to suit different regional characteristics or to address specific conservation issues and values. For example, selection of patch size thresholds might be related to average individual home range size or sustainable micro-population area of forest animals. Other criteria might be based on the spatial scale of natural regeneration and successional processes that provide continuity of forest ecosystems after disturbance. However, it is important to recognise that consistency of classes between sampling times is an essential component of any monitoring or comparative work.
For the evaluation of forest fragmentation based on global datasets (mostly 1 km2 pixel size), a conservative analysis would assume that details of forest patches smaller than 10 km2 cannot be analysed reliably because of the high impact of variation in shape and sub-pixel level structure. It can also be accepted that all forest patches larger than 300 km2 (30,000 ha) can be regarded as continuous forest, and thus make up the largest patch size class.
In the present analysis (Fig. 2), and potentially for future broad scale monitoring and data integration activities, logarithmic type scales are used, with the size class intervals delineated by rounded values in km2. (This minor distortion relative to a truly logarithmic scale will not affect the majority of statistical tests that could be applied in comparing patch size distributions). When higher resolution data are available, the scale might be extended to split the smaller patch size classes, while still keeping the established pattern of scaling. Equally, the upper end of scale could be expanded for broad scale or global level studies.
For the purpose of developing an integrated index of fragmentation, numeric ranks ranging from 1 (1-10 km2 patch) to 10 (> 300 km2) were assigned to classes, and each 1 km2 cell classed as forest was assigned to one of these ranks according to the size of patch it belongs to. The result (Fig. 2) is a visualisation of where forest occurs in large patches and where it occurs in small ones that may be clearer to non-experts than simple maps of forest cover. Such an analysis can also generate a statistical distribution of forest area among patch size (Ps) classes that can be used as a baseline for assessment and monitoring of forest condition in relation to the capacity for biodiversity preservation (Fig 3; see section on presentation issues below).
A limitation of the patch size analysis is that it tends to identify barely connected and/or irregularly shaped patches of forest as belonging to larger size classes than may be appropriate in terms of their ecological function. For example, [A] and [B] in figure 2 indicate forest patches of comparable size that are classed differently because of the presence or absence of small connections (of the order of the resolution of the raster data) to larger patches. Given the characteristics of coarse resolution data and the implications of connections or breaks between forest patches at this scale, additional components are needed to improve the strength of the evaluation. Furthermore, the patch size analysis alone does a poor job of distinguishing between the capacities to support biodiversity of outlying narrow branches of patches and core areas of forest patches within the same size class.
Figure 3. Statistical distribution of the forest area of Paraguay among different patch size classes. Such a distribution is an initial assessment of forest condition with respect to fragmentation and can be used as a baseline for monitoring purposes. If fragmentation of the forest increases, the amount of forest in the largest patch sizes will decrease and that in the smaller size classes will increase. However, it is common for deforestation to cause the disappearance of the smallest patches and a consequent reduction in the importance of that size class. Distributions can be compared between assessment times by using non-parametric statistical tests.
The effect of the interface between forest and non-forest (edge effects) can be addressed through the relatively commonly employed shape indices, such as perimeter to area ratios and edge-to-core ratios. However, these measures are more appropriate to ground level studies or high resolution data sets in which the forest cover is real and the extent of edge influence is well understood. For coarse resolution data in which patches may or may not represent actually contiguous forest, an alternative approach is to evaluate the percentage of the neighbouring cells that contain forest within a given radius of each cell (Spatially Weighted Forest Cover Density – SF). The radius can be chosen in the light of known scaling issues and concerns about specific ecological phenomena. In the current illustrative example (Fig. 4) a radius of 5 km was chosen as being consistent with the spatial accuracy of the data.
As can be seen in Figure 4,the value of SF is high when the sample cell belongs to the interior of, or is near a dense or solid patch that is comparable in size or larger than the radius used for calculation of SF. For areas at the periphery of, or distant from large continuous forest patches, SF is lower and mostly dependent upon patch shape and isolation. Small isolated patches have small amounts of forest in their immediate neighbourhoods ([A] in Fig. 4); the extreme case is the single-pixel forest patch that is separated by more than the SF evaluation radius from any other forest. Points along patch edges ([B]), and in patches of complex shape or in narrow strips of forest, where edge effects may influence forest status, also have low proportions of forest cover in their neighbourhoods, while points within larger patches [C] or continuous forest are entirely surrounded by forest. Therefore, spatially weighted forest density, SF provides a basis for expressing the role of edge effects, without dependence on accurate definition of the edge or limit of a given patch.
Like the patch size analysis, this analytical approach can be presented in a way that provides a clear spatially referenced visualisation (Fig. 4). This shows where forests are subject to edge and isolation effects as distinct from forest that is both part of a large patch (that is predominantly forested) and distant from the interface between forest and non-forest. Although the implications of these effects for forest biodiversity are strongly dependent on forest type and location, change in the amount and location of forest subject to such influences is likely to result in changing biodiversity status. Therefore, establishment of both visual and statistical baselines for this parameter is an essential step in the monitoring of forest biodiversity status.
Spatially weighted forest cover density, SF, is measured directly in percentage units. In the present example, (Fig 4 and 5), the scale is broken into ten equal intervals and the lower percentages are assigned to lower ranks. This reflects an implicit assumption that forest habitat that is less subject to the influences of edges and isolation is of greater value for forest biodiversity. Although this assumption may be invalid in some cases, it is a realistic view at the global scale. As with the patch size distribution, appropriate breaks between classes can be selected on theoretical and empirical grounds, and a statistical summary of the forest area in each class can provide a basis for assessment and monitoring of forest condition in this context (Fig. 5). However, consistency in approach between assessments is a critical component of any monitoring or comparative analysis. The use of GIS tools that retain the source data and intermediate parameters in separate grids, provides flexibility and ensures that a consistent approach can be applied across all data sets in a time series.
Figure 5. Statistical distribution of Paraguay’s forest area among different classes of spatially weighted forest cover density, SF, forest occurrence within a 5 km radius. Such a distribution is an initial assessment of forest condition with respect to fragmentation, which reflects proximity to non-forested areas and edges, and can be used as a baseline for monitoring purposes. If fragmentation of the forest increases, the amount of forest in the highest classes (i.e. those forest cells that are completely surrounded by forest) will decrease, and that in the smaller classes will increase. Distributions can be compared between times using non-parametric statistical tests.
A simplified version of the forest cover density analysis can be used to identify forest density zones, generalised outlines of the forest areas that are likely to have the highest integrity, those of least integrity and the intermediate values (Fig. 6). This approach provides a useful way of defining “core” forest areas (see below). It also provides a means of focussing attention on the forest areas of intermediate ‘quality’ or integrity. These are the areas that may appear least distinct to the non-expert observer, but are most likely to be immediately affected by policy and management decisions. Separate focus on this zone of intermediate forest density can enhance the visibility of pattern and change in the statistical data (see section on presentation below).
Quantification of the third component the effects of fragmentation, isolation, requires some measure of distance to other forest areas. However, the degree of isolation and/or the positive effects of interconnection are also dependent on the characteristics of the neighbouring forest. It is also true that many forest species are unable and/or reluctant to cross areas without forest cover (Laurance et al.,. 1997b), so forest areas that are directly connected to other forest areas are likely to be of greater value and more accessible to a greater range of forest species. The possibilities of dispersal to and from a particular forest area depend on the species of interest and the forest stand characteristics (among other factors).
Furthermore, the overall sustainability of biological systems is increased by the presence of relatively intact “core” areas surrounded by peripheral areas that are important in buffering the system as a whole against external impacts (see Fig. 6). These peripheral areas, in their turn, also benefit from connection to “core” areas, which provide necessary genetic resources (via animal migration or plant dispersal) that can be key to maintaining ecosystem function after natural or anthropogenic disturbance. Therefore, forest patches that are connected by forest to core forest areas may be viewed as more sustainable and of higher biological value that areas of similar forest cover density and overall shape that are not connected to core forest.
Core forest area, and connection to it, may be defined using thresholds appropriate to the particular components of biodiversity of interest, management considerations and properties of individual forest types. In the present study, for general illustration core forest area was defined using two criteria that are appropriate to the coarse scale of the data (Fig. 6):
1. Core area is represented by continuous forest with density (SF ) more than 90%
2. The size of an individual core area must be at least 100 km2
The distance to core forest areas, via cells containing forest cover, estimates the degree of connection of forest cells to core areas. The connectivity, CF, is inversely proportional to distance from core areas, ranging from 10 (core area) to 1 (24-27 km), and in this example, forest cells at distances greater than a threshold of 27 km from core forest are regarded as effectively isolated and assigned a CF value of 0. Forest cells that are connected to core forest by between one and 27 km of forest are assigned to intermediate classes of connection (Figure 7).
Like the previous two indicators, this approach can be used to provide both a spatially referenced visualisation and a statistical summary (Figure 8) of which forest areas are likely to be in the best condition for preserving forest species. In this case, the best condition refers to the closest connection to core forest, and therefore the greatest accessibility to the greatest range of forest species. Although this approach is proposed for the analysis of relatively low resolution (1 km) data, where individual pixels may include mosaic patterns, and species diversity and successional stage may not be incorporated, the same principles would apply at more detailed scales.
Figure 8. Statistical distribution of forest area in Paraguay among different classes of connectivity, CF. This distribution could serve as a baseline for monitoring the changes in forest capacity to retain biodiversity. As forest fragmentation increases, the amount of core and highly connected forest will decrease, and unconnected or remotely connected forest area will increase both absolutely and as a proportion of the total. Some scenarios of forest regeneration could result in increasing amounts of connected forest.
Despite the individual limitations of each of the above indices, between them they cover all three important aspects of forest fragmentation effects and make it possible to quantify most variations in the spatial distribution of forest cover. Patch size, Ps, facilitates comparison of forest stands larger than some minimal patch size threshold and below a size that can be regarded as continuous forest, regardless of variation. Spatially weighted forest cover density, SF, provides a way of identifying both small, dispersed patches and areas subject to edge effects. However, it provides little detail at intermediate values and gives insufficient information about the interconnection of forest patches with forest patches of different status. The connectivity index, CF, permits ranking patches of similar size in relation to their probable accessibility to forest species, and provides a way of distinguishing among forest areas of intermediate sizes and densities.
As all three indices range between 0 and 10, it is feasible to integrate them, using averaging or some more complicated method of combination, to provide a combined measure that is similarly easy to understand and interpret. The contributions of the individual indices can be adjusted using numeric coefficients to reflect the conceptual weight attached to the different factors:
FF = x(PS) + y(SF) + z(CF)
As patch size and forest density are interrelated, while connectivity is conceptually distinct and potentially very important in sustaining forest biodiversity, we suggest the following composite index of forest spatial integrity as a starting point for analysis of forest condition:
FF = 0.25PS + 0.25SF+ 0.5CF
Other coefficients could be adopted to reflect a different focus.
The example application of this forest spatial integrity index, FF, to the forest cover data for Paraguay (Fig. 9) demonstrates the visual impact and clarity provided by this approach, which could ensure its utility for decision making. The statistical distribution of forest area among the different integrity index classes (Fig. 10) provides a baseline for monitoring forest spatial integrity over time. Such monitoring is an essential component of evaluating policy effectiveness and the impact of management decisions.
Great care must be taken in the presentation of results of such assessment and monitoring. Data in a mapped context may be the most useful for supporting site-specific decision-making and, conceivably, for scenario testing. Data in statistical form are potentially more useful for monitoring and evaluation of policy effectiveness. It is crucial that consistent methods are applied for comparisons in space and time and that original data and analyses are retained to permit reanalysis in the event that changes in thresholds or approaches are deemed appropriate.
Statistical data need to be presented in absolute areas rather than as percentages. Changes in percentages of forest cover in lower integrity categories may reflect either an improvement in the integrity of formerly low value areas, or simply a loss of forest area from those categories without an increase in others.
Figure 10. Statistical distribution of the forest area of Paraguay among different classes of spatial integrity as expressed by the Index of Forest Spatial Integrity, FF, which combines information on patch size, shape and isolation. Such a distribution can be used as a baseline for monitoring forest condition with respect to fragmentation. If deforestation reduces forest area uniformly, the totals will decrease uniformly across the distribution. The loss of small isolated forest remnants or irregular patches will be reflected in a reduction in the area in the lower integrity classes, while the loss of forest with high integrity can be detected from a loss of forest area with high integrity value. Distributions can be compared between assessment times by using non-parametric statistical tests.
Added clarity of both mapped and statistical presentation of data on forest spatial integrity may be obtained by delineating the three different forest density zones (Fig. 11) as derived from the simplified presentation of the Spatially Weighted Forest Density (Fig. 6). This presentation might help the non-expert user to visualise the spatial relations among forest patches in relation to their overall spatial integrity index and to anticipate the effects of changes in the landscape more vividly. It also makes it possible to view the statistical distribution among intermediate integrity classes in more detail (Fig. 12), as the scale can be expanded by excluding large areas of high and low integrity forest from the presentation.
Figure 12. Statistical distribution of the forest area of Paraguay, according to the three component indices and the integrated Index of Forest Spatial Integrity. These graphs include only the forest falling within the zone of intermediate density (cf. Fig 6) and permit greater scrutiny of the distribution of forest among intermediate classes of each index and, potentially its change over time, than Figure 10. Forest areas of intermediate integrity are those most likely to be affected by changes in policy and management, and therefore such a form of presentation may be of use to inform decision-makers.
Another application, to data from the Ukraine (Figs 13 and 14), shows that the analysis produces visually consistent results and that it is feasible to apply this approach widely. It shows clearly that the Ukraine has smaller amounts of core and high integrity forest, but retains some areas of intermediate integrity that can be identified and prioritised in policies to promote forest conservation and sustainable use. This analytical approach could be implemented in a global assessment of forest fragmentation to provide a baseline for monitoring change in forest cover and its integrity and provide insight into the changing capacity of the world’s forests to retain their biodiversity.
Spatial integrity is not in itself a sufficient measure of forest capacity to maintain biodiversity. Of the other influences that are important, a key factor is the influence of human population and activity on remaining forest areas. This is discussed in the following section.
Figure 14. Statistical distribution of the forest area of Ukraine among different classes of spatial integrity as expressed by the Index of Forest Spatial Integrity, FF, which combines information on patch size, shape and isolation. Such a distribution can be used as a baseline for monitoring forest condition with respect to fragmentation or for comparison with data from other locations (e.g. Fig 11). In (a) the distribution is shown for all the forest area of the Ukraine, while in (b) only forest falling within the zone of intermediate density (see Fig. 13) is included, allowing more detailed scrutiny of the distribution among intermediate integrity classes.