Abstract
Habitat fragmentation, characterized by changes over time in the composition and spatial configuration of critical early winter habitat for woodland caribou (Rangifer tarandus caribou), may be occurring near Revelstoke, British Columbia as a result of timber harvesting and natural disturbances such as wildfires. The methodology employed to test this hypothesis is based on the integration of remote sensing, Geographical Information Systems (GIS), and landscape models. Habitat suitability models were created to represent the landscape for the study area in a past (1975) and more recent condition (1997). Elevation, slope and forest stand age data for each time period were used to develop the habitat models. Land cover information used to build the habitat suitability models was generated using Landsat Multispectral Scanner (MSS) imagery for 1975 and Thematic Mapper (TM) imagery for 1997, and a Hybrid Decision Tree Classifier.

Based on a comparative analysis of selected spatial landscape metrics calculated for each time period, changes in the composition and spatial configuration of early winter habitat were quantified. Overall, the amount of suitable winter habitat available in 1997 represented a decrease of approximately 832.88 hectares (8.53%) from 1975 levels. As a result of the observed disturbance pattern, early winter habitat patches in 1997 were smaller and more uniform in size than in 1975, as suggested by a decrease in mean patch size and patch size coefficient of variation indices between time periods. Spatial pattern indices calculated for habitat patches in 1997 also indicated a reduction in geometric complexity, interior core area, and mean proximity, while patch abundance and density, edge density, and juxtaposition had increased since 1975. The habitat fragmentation results are consistent with those reported in the literature in other regions, and indicate that fragmentation of early winter caribou habitat is occurring in the study area.

Keywords
Habitat Fragmentation, integrated model, landscape metrics, temporal analysis, hybrid classifier, caribou, disturbances, environmental monitoring, British Columbia.

The region to the north of Mount Revelstoke and Glacier National Parks, near Revelstoke, British Columbia, represents a landscape under change. The influence of extensive logging and natural disturbances such as wildfires in the region has resulted in the replacement of large areas of the natural forest cover with an extensive pattern of clear cuts and burns. As a result, suitable habitat patches for local woodland caribou (Rangifer tarandus caribou) and natural corridors allowing connectivity between these sites are possibly being fragmented.

Habitat fragmentation is a landscape evolution process characterized by both the reduction in the total amount of suitable habitat available (habitat loss) and the spatial isolation of the remnant habitat patches over time (McGarigal and Marks 1995). Biological conservation theory recognizes that habitat fragmentation may eventually result in the separation of vertebrate populations into small island sub-populations that are more prone to local extinction. Therefore, to ensure the long-term survival of woodland caribou in the Revelstoke region, extractive resource activities and other environmental disturbance factors should be managed to maintain regional habitat connectivity. The development of an appropriate method is required to quantify and assess the spatial characteristics of environmental disturbances and facilitate landscape-scale monitoring of caribou habitat fragmentation over time.
 

The main goal of this research is to determine the spatial effects of timber harvesting and wildfires on caribou habitat composition and configuration in the Revelstoke forest region for the period from 1975 to 1997. The research hypothesis being tested is that fragmentation of critical early winter caribou habitat is occurring in the study area. Two interconnected objectives are required to address the main hypothesis.

1.  Habitat suitability model development for 1975 and 1997
2.  Quantification and assessment of habitat fragmentation. This was evaluated in terms of:

• habitat loss as reflected by changes in overall habitat composition and
• changes in the spatial characteristics and configuration of early winter habitat occurring between 1975 and 1997.

The decline of woodland and mountain caribou populations in the last century as a result of human settlement has prompted scientific research on the habitat requirements of the species, the spatial distribution of critical habitat areas, and the effect of human and natural disturbances on caribou habitat distribution and connectivity. As part of this initiative, several researchers have identified that early winter habitat of mature and old growth cedar-hemlock forests may constitute the most critical component for continued survival of mountain caribou (Rominger and Oldemeyer 1989). McLellan et al. (1994) argued that emphasis should be placed on conserving early winter habitat based on potential conflict with timber harvesting activities within the Revelstoke forest region. Several studies concerning woodland caribou (Bradshaw et al. 1997) have illustrated the utility of remote sensing, GIS, and landscape ecology techniques for caribou habitat analysis.

Remote sensing/GIS integration
Recent research on remote sensing/GIS integration supports a combination of data sources and techniques to provide a more comprehensive analysis of environmental change (Wilkinson 1996). Recent classification algorithms applied in remote sensing which take advantage of these integrated techniques include knowledge-based classifiers. Decision tree classifiers represent a form of knowledge-based classification that provide important advantages over conventional methods. For example, the Hybrid Decision Tree algorithm described by Friedl and Brodley (1997) allows the incorporation of different data types and classification algorithms in a single approach.

Landscape ecology
Landscape ecology research supports the application of spatial pattern metrics for quantifying environmental change. Several studies investigating the effects of human disturbances such as timber harvesting on forested environments (Reed et al. 1996; Sachs et al. 1998) have shown significant relationships between changes in landscape metrics and underlying ecological processes. A sample of landscape metrics commonly used to quantify the effects of forest fragmentation is evident within the literature. Specifically, these include measures quantifying mean patch size, interior core area, inter-patch distance, edge density, and mean patch shape. This selection of metrics was chosen in this research in order to measure the amount of change and fragmentation in suitable early winter caribou habitat patches between 1975 and 1997. This study represents the first attempt to execute a long-term temporal analysis of caribou habitat change in the study area using a combination of spatial methodologies.
 

The suitability or value in terms of caribou habitat quality was assessed for each spatial feature within the study area for each time period. One common approach for quantitative evaluation of habitat is the development of a Habitat Suitability Index (HSI) model (U.S. Fish and Wildlife Service 1981; Rickers et al. 1995) to establish the value of a particular habitat based on a species’ observed preference or use of different land cover types. An HSI model has been developed for mountain caribou in the study area by McLellan et al. (1995). Based on their work, the HSI model in our research utilizes elevation, slope, land cover or habitat unit type, and stand age:

(Notation 1)           HSI_total = (HSI_elevation * HSI_slope * HSI_{habitat unit} * HSI_{stand age})^1/4

The total or final HSI value for any location within the study area represents the geometric mean of HSI values for each habitat variable. The elevation and slope variables were derived from a DEM for the study area. Stand ages for 1997 were projected from forest inventory stand origin data. The remaining HSI variable (Habitat Unit) for 1997 was derived from a decision tree classification algorithm. (for a more detailed discussion see Deuling 1999)

Hybrid decision tree classification (1997 model)
The classification procedure used to map habitat units for 1997 involved the integration of GIS and remote sensing analysis in a Hybrid Decision Tree (HDT) algorithm. This type of approach involves the recursive partitioning of a data set into smaller subdivisions or classes with a different classification algorithm being applied at each split in the tree structure. The Decision Tree structure developed for this research applied four general components to separate spatial features (contiguous pixels of the same cover type) into the appropriate class designation. These classes included recent cuts and burns and classes related to land cover and vegetation. GIS attribute selection, brightness differencing, K-Means unsupervised classification, Maximum Likelihood Classification (MLC) and spatial and contextual decision rules were all incorporated into the HDT algorithm. The end result of this procedure was a map representing habitat units in 1997. The habitat unit map was then reclassified into the appropriate early winter habitat suitability indices. The geometric mean of the four individual HSI variable images (elevation, slope, habitat unit, and stand age) was then calculated, using Notation 1, to create a final image of early winter habitat suitability for the 1997 time interval.

Classification and Mapping of Habitat Variables (1975 model)
A 1975 habitat unit map, which could be compared to the 1997 map, was derived for this analysis using a method similar to that described by Reed et al. (1996). The 1997 habitat unit map produced for the study area was used as a base dataset from which to project changes in classes back into the past and recreate a 1975 habitat and disturbance map that could be directly compared to the present landscape structure. A second Hybrid Decision Tree Algorithm was implemented to classify the 1975 Landsat Multispectral Scanner (MSS) satellite image into the same habitat and disturbance units of the 1997 image. The class designation of each pixel before and after stand disturbance was compared and a series of IF/THEN statements was applied to identify certain types of forest class transitions occurring between 1975 and 1997. The geometric mean of the four individual HSI variable images (elevation, slope, habitat unit and stand age) was then calculated to create a final image of early winter habitat suitability for the 1975 time interval (Deuling 1999).

Calculation and Comparison of Spatial Pattern Metrics
The landscape boundary defining the region of analysis corresponded to the home range of the local sub-population of mountain caribou and remained constant between time periods so that landscapes of similar extent were being compared. The past (1975) and most recent (1997) landscapes were constructed in such a way to represent a mosaic of habitat patches that differed based on the early winter habitat suitability rating. The original ratio range of HSI values (between 0 and 1) was reclassified into 10 ordinal rank HSI classes to create discrete habitat suitability classes for fragmentation analysis, with the lowest habitat suitability represented by HSI class 1 and the highest or most suitable early winter habitat as HSI class 10. Spatial metrics were calculated for each of the 10 ordinal habitat suitability classes using the Patch Analyst software extension for the ArcView GIS program (Elkie et al. 1999). Of primary interest were the directional changes in spatial metrics for the high suitability classes (HSI classes 8, 9, 10) between 1975 and 1997. Therefore, the focus of the fragmentation analysis was on changes in HSI class level indices rather than a landscape level analysis.
 

Temporal changes in habitat composition
The most extensive areas of suitable early winter habitat in 1975 occurred in the Carnes Creek and Downie Creek watersheds (see Figure 1). A visual comparison of these areas between 1975 and 1997 reveals that large portions of suitable early winter habitat (HSI classes 8-10) have transformed from a high HSI class in 1975 to a lower HSI class in 1997. In particular, in the central portion of the Downie Creek valley (Figures 1a & 1b) and in the southern slopes of the Carnes Creek valley (Figures 1c & 1d), the spatial extent of HSI classes 8, 9, and 10 and the spatial characteristics and configuration of these important habitat areas has been affected by disturbance. Large, contiguous areas of suitable habitat patches in 1975 have been reduced in size by the conversion to low quality habitat. The grey patches occurring in these areas in 1997 correspond primarily to recent timber harvest cuts. Based on these observations, the HSI maps provided in Figure 1 suggest that timber harvesting and wildfires have affected the composition and configuration of quality early winter habitat in the study area between time periods.
 

The spatial metrics calculated for 1975 and 1997 habitat suitability maps are compared in Table 1. The areas with the highest habitat suitability rating (classes 8, 9, 10) represent the focal patch types of the fragmentation analysis. In order to focus the discussion, the results for the lower rated habitat classes are not presented. The first metric presented is Class Area and compares the overall change in area of each habitat suitability class between 1975 and 1997. If high quality habitat is considered to be that above 0.70 HSI, then 8.53% of the quality early winter habitat has been lost since 1975, indicating that forested areas providing high quality early winter habitat are being disturbed at a faster rate than replacement by stand regeneration. Those areas having habitat suitability ratings between 0.70 and 0.80 (HSI class 8) experienced the largest reduction in area (11.05%). Figure 2 shows the breakdown of the habitat loss by disturbance type and compares the relative effects of fire and timber harvesting on the observed habitat loss. Stand disturbance due to timber harvesting appears to be the primary cause of the net decline in area of high quality habitat.
 
 

Temporal changes in habitat configuration
The observed habitat loss or changes in habitat composition are accompanied by changes in spatial configuration of habitat patches. Table 1 indicates that patch abundance (Number of Patches) and Patch Density for each focal HSI class also increased as a result of landscape disturbances, as would be expected with more habitat patches in the same total area.

Mean Patch Size decreased for each of the classes of interest between 1975 and 1997, and this change was significant (p < 0.0891, two-way ANOVA) for HSI class 8, but not statistically significant for the other focal patch types (HSI class 9 or 10) (p > 0.10, two-way ANOVA). The significant decrease in patch size for HSI class 8 probably results from the disproportional removal of this habitat class by timber harvesting compared to the other habitat suitability categories. In contrast, areas falling within HSI classes 9 and 10 were less frequently targeted by harvest operations. However, the ecological significance of the observed decrease in mean patch size (as indicated by a change of -14.38% for HSI class 9 and –8.65% for HSI class 10) should not be underestimated.

Edge Density (the amount of edge relative to the area of all classes) increased the most for HSI class 9 (4.98% change) with only a slight increase for class 10 (0.95%). In contrast, this metric actually decreased for HSI class 8, a result that at first seemed to contradict the expected results of fragmentation. However, upon close visual inspection of the changes in the spatial distribution of HSI class 8 in 1975 compared to 1997, it appears that the main effect of harvesting and fires on this particular habitat class is patch attrition or disappearance. In this case, the amount of edge between forest and non-forest cover types actually decreases as low quality habitat patches coalesce over time.

The Mean Shape index in Table 1 shows a decrease between time periods for HSI classes 8 and 9, indicating that suitable habitat patches are becoming less geometrically complex over time, although the amount of change in this index is not large.

The Mean Proximity Index (MPI) decreased substantially for each of the focal HSI classes, suggesting that the isolation and degree of fragmentation for suitable early winter habitat patches has increased as a result of timber harvesting and fire activity between 1975 and 1997. The metric equation (see McGarigal and Marks 1995) for MPI is a product of both the average patch area of a class and the mean edge-to-edge distance between patches. Based on this formula, it follows that a larger MPI value indicates a class in which patches are distributed in larger, more contiguous areas that are located in closer proximity to patches of the same type (McGarigal and Marks 1995). Therefore, the decrease in MPI reported in Table 1 appears to confirm that areas of suitable habitat are being fragmented into smaller patches that are more dispersed or isolated from each other over time.

Mean Core Area is the average patch size remaining after removing an edge buffer of 100 metres representing the area influenced by the edge effects along patch boundaries compared to the interior or core conditions. The Mean Core Area decreased substantially for HSI classes 8 and 9 by 14.02 and 31.25% respectively. Only a slight decrease was reported for the highest suitability class (HSI class 10) with a 1.64% change from the 1975 Mean Core Area value.

The results of the habitat patch analysis reported here indicate that both the landscape composition and configuration have been altered over the 22 year period of the study as a result of timber harvesting and wildfires. The observed directional changes in the spatial metrics calculated for each time period are consistent with results for similar studies and appear to confirm the observed spatial effects of forest fragmentation recognized in the literature. Mean Patch Size and the Mean Proximity indices showed the largest change between 1975 and 1997 for the HSI classes of interest in the study area. This suggests a stage of landscape evolution that fits with both the fragmentation and shrinkage phases described by Forman (1995).
 

Caribou habitat suitability models, Landsat MSS, TM and a forest inventory GIS database were integrated to examine the spatial effects of timber harvesting and wildfires on a forested landscape near Revelstoke, British Columbia. It was hypothesized that fragmentation of early winter habitat critical to a local sub-population of mountain caribou had occurred in the study area as a result of these specific disturbance factors, and that these landscapes changes could be quantified and assessed with landscape metrics applied to the satellite image and GIS mapping products.

A disproportionate disturbance of old growth cedar hemlock stands by harvesting is suggested to be the leading cause of a net decline in the amount of high quality early winter habitat in the study area between 1975 and 1997. Changes is the spatial configuration of suitable habitat patches were also evident over the 22 year period of the study as patch abundance and density, edge density, and interspersion increased while mean patch size, patch size variation, patch shape, proximity, and core area all decreased between time periods. This supports the position that, as a result of timber harvesting and wildfires the prime early winter habitat for the caribou is becoming fragmented over time.

The value of this work can be recognized on a number of fronts. It represents further support for the need to more effectively manage resource extractive activities. It provides a methodology that can now be used for future environmental monitoring by park managers. Finally, it represents an example of the value of using integrated spatial models in a temporal analysis of environmental change.
 

The financial support of Parks Canada, the Natural Sciences and Engineering Research Council of Canada, and the University of Calgary is gratefully acknowledged. We are thankful for the assistance of M. Hindle, P. Deuling, and J. Hansen during field data collection and for the recommendations and advice of G. Gerylo.
 

Bradshaw CJA, Boutin S, Hebert DM. 1997. Effects of petroleum exploration on woodland caribou in northeastern Alberta. Journal of Wildlife Management 61(4): 1127-1133.

Deuling, M.J. 1999. Quantification and Assessment of Caribou Habitat Fragmentation: An Integrated RS, GIS, and Landscape Ecology Method, Unpublished M.Sc. Thesis, Department of Geography, University of Calgary

Elkie PC, Rempel RS, Carr AP. 1999. Patch Analyst User’s Manual: A Tool for Quantifying Landscape Structure. Ontario Ministry of Natural Resources NWST Technical Manual TM-001. Thunder Bay, Ontario.

Forman, RTT. 1995. Some general principles of landscape and regional ecology. Landscape Ecology 10(3):133-142.

Friedl MA, Brodley CE. 1997. Decision Tree Classification of land cover from remotely sensed data. Remote Sensing of Environment 61: 399-409.

McLellan B, Flaa J, Super M. 1995. Developing a Habitat Suitability Model for Mountain Caribou in the North Columbia Mountains: Preliminary Report # 4. Ministry of Forests, Revelstoke, B.C. 23 pp.

McLellan B, Flaa J, Super M. 1994. Habitats used by mountain caribou in the North Columbia Mountains: Preliminary Report # 2. Ministry of Forests, Revelstoke, B.C. 15 pp.

McGarigal K, Marks BJ. 1995. FRAGSTATS: Spatial pattern analysis program for quantifying landscape structure. USDA Forest Service General Technical Report PNW-GTR-351. Cornvallis, Oregon.

Reed RA, Johnson-Barnard J, Baker WL. 1996. Fragmentation of a Forested Rocky Mountain Landscape. 1950-1993. Biological Conservation 75: 267-277.

Rickers JR, Queen LP, Arthaud GJ. 1995. A proximity-based approach to assessing habitat.  Landscape Ecology 10(5): 309-321.

Rominger EM, Oldemeyer JL. 1989. Early-winter habitat of woodland caribou, Selkirk Mountains, British Columbia. Journal of Wildlife Management 53(1): 238-243.

US Wildlife Service. 1981. Standards for the development of suitability index models. Ecological Services Manual 103. US Department of Interior, Fish and Wildlife Service, Division of Ecology Services, Government Printing Office, Washington,D.C. 68pp. + appendices.

Wiens JA. 1976. Population response to patchy environments. Annuls of Rev. Ecological Systems 7: 81-129.

Wilkinson GG. 1996. A review of current issues in the integration of GIS and remote sensing data. International Journal of Geographical Information Systems 10(1): 85 - 101.

Authors
Medina J. Deuling, GIS/RS Technologist, Department of Geography
University of Calgary, 2500 University Drive NW, Calgary, Alberta, Canada T2N 1N4.
Email: deuling@ucalgary.ca, Tel: 403-220-5588, Fax:403-282-6561

Clarence G. Woudsma, Ph.D., Assistant Professor, Department of Geography
University of Calgary, 2500 University Drive NW, Calgary, Alberta, Canada T2N 1N4.
Email: cwoudsma@ucalgary.ca, Tel: 403-220-2576, Fax:403-282-6561

Steven E. Franklin, Ph.D., Professor, Department of Geography
University of Calgary, 2500 University Drive NW, Calgary, Alberta, Canada T2N 1N4.
Email: franklin@ucalgary.ca, Tel: 403-220-6192, Fax:403-282-6561