Home | Archives | About | Login | Submissions | Notify | Contact | Search

 ES Home > Vol. 6, No. 2 > Art. 16

Copyright © 2002 by the author(s). Published here under license by The Resilience Alliance.
Go to the pdf version of this article.

The following is the established format for referencing this article:
Poulin, M., D. Careau, L. Rochefort, and A. Desrochers. 2002. From satellite imagery to peatland vegetation diversity: how reliable are habitat maps? Conservation Ecology 6(2): 16. [online] URL: http://www.consecol.org/vol6/iss2/art16/


Report

From Satellite Imagery to Peatland Vegetation Diversity: How Reliable Are Habitat Maps?

Monique Poulin1, Denis Careau2, Line Rochefort1, and André Desrochers1


1Université Laval2Société de Mathématiques appliquées


ABSTRACT

Although satellite imagery is becoming a basic component of the work of ecologists and conservationists, its potential and reliability are still relatively unknown for a large number of ecosystems. Using Landsat 7/ETM+ (Enhanced Thematic Mapper Plus) data, we tested the accuracy of two types of supervised classifications for mapping 13 peatland habitats in southern Quebec, Canada. Before classifying peatland habitats, we applied a mask procedure that revealed 629 peatlands covering a total of 18,103 ha; 26% of them were larger than 20 ha. We applied both a simple maximum likelihood (ML) function and a weighted maximum likelihood (WML) function that took into account the proportion of each habitat class within each peatland when classifying the habitats on the image. By validating 626 Global Positioning System locations within 92 peatlands, we showed that both classification procedures provided an accurate representation of the 13 peatland habitat classes. For all habitat classes except lawn with pools, the predominant classified habitat within 45 m of the center of the validation location was of the same type as the one observed in the field. There were differences in the performance of the two classification procedures: ML was a better tool for mapping rare habitats, whereas WML favored the most common habitats. Based on ordinations, peatland habitat classes were as effective as environmental variables such as humidity indicators and water chemistry components at explaining the distribution of plant species and performed 1.6 times better when it came to accounting for vegetation structure patterns. Peatland habitats with pools had the most distinct plant assemblages, and the habitats dominated by herbs were moderately distinct from those characterized by ericaceous shrubs. Habitats dominated by herbs were the most variable in terms of plant species assemblages. Because peatlands are economically valuable wetlands, the maps resulting from the new classification procedure presented here will provide useful information for land managers and conservationists.

KEY WORDS: Landsat 7, coarse filter, error matrix, habitat distinctiveness, habitat variability, mask procedure, maximum likelihood classification, peatland habitats, plant species assemblages, remote sensing, supervised classification, wetland conservation.

Published: December 9, 2002


INTRODUCTION

Concerns about the loss of biological diversity have encouraged ecologists to improve their ability to trace species distributions and occurrences over large spatial scales and short periods of time (Schoch and Dethier 1996, Boone and Krohn 1999, Pearce and Ferrier 2001). Sampling methods based on the coarse filter approach (Noss 1987, 1996) have gained popularity as efficient tools for protecting a large part of species diversity (Franklin 1993, Wilcove 1993, Schwartz 1999, Hughes et al. 2000). Although conservation strategies based on habitats or plant communities have their limitations, e.g., plant communities may not always be a good surrogate for the distribution patterns of rare species (Panzer and Schwartz 1998), they remain an essential component of land use planning and reserve selection procedures (Pressey 1994, Margules and Pressey 2000).

Remote sensing offers promising tools for detecting and mapping regional landscape patterns and processes (Roughgarden et al. 1991, Kasischke et al. 1997, Joint and Groom 2000). For these purposes, satellite imagery presents several advantages over other remote sensing techniques such as the interpretation of aerial photographs. Satellite imagery covers larger areas, has a greater spectral resolution, is already in digital format, and is processed more homogeneously across a whole region (e.g., in Fig. 1) and at a lower cost (Fuller et al. 1989, Konrad and Rempel 1990, Holopainen and Wang 1998, Mumby et al. 1999) than alternative methods. So far, satellite image classification has been developed for a wide variety of habitats, including agricultural lands (Oetter et al. 2001), forest stands (Rey-Benayas and Pope 1995), grasslands (Lauver and Whistler 1993), and wetlands (see references in following paragraph). Nevertheless, satellite imagery is not often used to investigate peatlands, despite the fact that they are dominant, regionally threatened landscape elements in many parts of the northern hemisphere (Gorham 1990, Lappalainen 1996).

Previous work that attempted to classify different wetland habitats was restricted mainly to broadly defined habitat types and carried out for specific purposes such as quantifying the extent of potential foraging sites for wildfowl (Hogson et al. 1987, Herr and Queen 1993, Grenier et al. 1994, Tatu et al. 1998). In peatlands, satellite imagery has been used to detect water flow (Glaser et al. 1981, Glaser 1983) and investigated with regard to its ability to map vegetation types (Glaser 1989). However, only a few studies have successfully mapped several types of peatland habitat (Palylyk et al. 1987, Quarmby et al. 1997, Boresjö Bronge and Nöslund-Landenmark 1999, Aaviksoo et al. 2000). Although only one of these four studies presented detailed quantitative information about the accuracy of their classifications, they indicate the potential of remote sensing for mapping peatland habitats defined using a high resolution.

With the aim of conserving species diversity, there is a clear need to improve our ability to distinguish one habitat type from other similar types. Moreover, if we are to adopt the coarse filter approach, we also need to be able to establish reliable links between certain habitat types and the occurrence of particular species (Roughgarden et al. 1991, Stoms and Estes 1993, Lewis 1998). Methods have been developed for mapping habitats based on spectral reflectance patterns and establishing their correspondence with the occupancy patterns of plant or animal species (Scott et al. 1993, Lauver 1997, Nøhr and Jørgensen 1997, Fuller et al. 1998, Debinski et al. 1999, Nagendra and Gadgil 1999). Nevertheless, the few remote sensing studies that did address the distribution of plant species in wetlands dealt mainly with pure vegetation stands and/or were conducted 1–2 m above the canopy with portable field spectrometers (Budd and Milton 1982, Peñuelas et al. 1993, Zhang et al. 1997, Spanglet et al. 1998). For conservation purposes, these advances should be applied to broader areas so that data collected from satellite-borne sensors can be used to predict the occurrence of wetland species and vegetation structure patterns across the landscape.

In this paper, we assess the use of Landsat 7/ETM+ (Enhanced Thematic Mapper Plus) satellite imagery for mapping peatland habitats at spatial scales relevant to regional (1000 km2) diversity management. More specifically, we compare the accuracy of two supervised classification procedures for mapping 13 habitat classes that we defined a priori based mainly on vegetation structure. We also assess how much variability in the vegetation structure and plant species patterns can be explained by our classification system vs. 15 environmental variables, and how distinct and variable the species compositions are in the 13 habitats defined a priori.


STUDY AREA

The study area consists of about 5000 km2 of lowlands located on the south shore of the St. Lawrence River in the province of Quebec, Canada (Fig. 2). It is dominated by marine sands whose maximal altitude is < 150 m above sea level. In this region, peatlands occur on poorly drained terrain that is normally found between thin, littoral strings of sand and gravel. The region is characterized by forested (45%) and agricultural (40%) lands (Robitaille and Saucier 1998). Peatlands cover about 4% of this region; they are ombrotrophic or weakly minerotrophic and therefore dominated by Sphagnum mosses.


METHODS

Selection of satellite image

We purchased a Landsat 7/ETM+ (Enhanced Thematic Mapper Plus) scene taken on 14 July 1999, a period that coincided with the full development of most of the vegetation types in our study area (see http://landsat7.usgs.gov/). The image contains radiometric information recorded by a scanner in six spectrally defined channels of 30 x 30 m pixels as well as two additional channels, one for thermal infrared radiation measuring 60 x 60 m pixels and one panchromatic channel of 15 x 15 m pixels (Jones and LeAnn 2000). The panchromatic and thermal bands were not used for the classification work because of their low spectral and spatial resolutions, respectively.

We proceeded with a single-date imagery classification because of logistical and financial constraints. However, we do not think that this impaired our ability to discriminate among habitat classes, because there are no pronounced differences in plant phenology in peatland habitats: most shrub species are evergreen, and herb development is quite synchronous among species. Moreover, the image was taken at mid-summer, i.e. when the water table was likely to be below its maximum (Price 1997, Verry 1997, Van Seters and Price 2001), which made it easier to discriminate among habitats.

Habitat classes and training sites

We mapped peatland habitats using a supervised classification approach (Richards and Jia 1999) that requires habitat classes to be defined a priori. The spectral signature (radiance) of a given habitat class was then determined by sampling pixels known to belong to that class, with the help of "training sites." Finally, a classification algorithm assigned each pixel of the image to one of the habitat classes based on the statistical properties of the radiance data of each habitat class. Following Palylyk et al. (1987) and relying on the ability of three experts to recognize habitat patterns on aerial photographs, we defined 17 habitat classes based on the structural properties of the vegetation as well as on species discrimination among some forested habitat classes. Four of the habitat classes originally defined were eliminated in the early stages of the study because they were rare or hard to access, leaving 13 habitat classes in the final classification (Table 1, Appendix 1). We restricted our study to peatlands that were less than 50% covered by trees > 4 m high. Cedar (Thuja occidentalis L.) and maple (Acer spp.) forests on peat were not considered, nor were the lagg zones that occur at the edges of peatlands and therefore represent an ecotone.

Training sites were located in 15 different peatlands that ranged in size from 60 to 756 ha and that were chosen to cover the entire range of the 13 habitat classes that we had defined a priori. These sites were selected with the help of aerial photographs, followed by ground-truthing conducted in the fall of 1999. The minimum area for a training site was 0.5 ha, i.e., about six pixels. The geographical location of each training site was determined using a Differential Global Positioning System (DGPS) receiver. Following the ground-truthing of these training sites, 1021 pixels were sampled on the nongeoreferenced satellite image for an average (± SD) of 79 ± 42 pixels per habitat class. This number of training sites was considered sufficient because only one class (LawnPool) had fewer than 50 (ca. 28) sampled pixels (Richards and Jia 1999). For this step, DGPS locations of training sites were positioned on the georeferenced satellite image using ArcView GIS software (Environmental Systems Research Institute 1996). Information from this georeferenced image was transposed visually onto the nongeoreferenced satellite image. We are confident that this method did not bring any bias into our sampling of training sites, all of which were located within large homogeneous zones that were easily recognizable on both images. Pixels corresponding to the training sites were not sampled on the georeferenced image to avoid spatial distortion of individual pixels and keep the proportions of each class on the image unchanged (cf. Richards and Jia 1999).

Distinguishing peatlands from the matrix

Before classifying peatland habitats, we built a mask to isolate the pixels corresponding to peatland habitats from the rest of the image. Each pixel of the image was thus assigned to one of two superclasses: peatland or nonpeatland. The spectral signature of the peatland superclass was based on the training sites sampled in 11 of our 13 peatland habitat classes (Table 1). Habitats 1 (spruce forest with open canopy) and 12 (spruce thickets with pools) were excluded from the peatland superclass because they tended to overestimate the surface area of peatlands. These two habitat classes were nevertheless considered when classifying peatland habitats thereafter and are thus part of the resulting habitat maps.

The nonpeatland superclass was composed of two classes: a hydrographic class corresponding to rivers and lakes, and a generic class corresponding mainly to forests, urban areas, and agricultural lands. The spectral signature of the hydrographic class was determined by sampling pixels from water bodies located in all parts of the image, whereas the spectral signature of the generic class was determined by sampling 1000 pixels randomly over the entire image. Although this procedure could capture peatland pixels, their contribution to the statistical properties of the spectral signature of the generic class was likely to be small compared to the total extent of other land cover types.

Each pixel of the image was attributed to either the peatland or the nonpeatland superclass according to normal maximum likelihood functions, which were calculated to discriminate between the two superclasses. When the ratio between the functions for the peatland and nonpeatland superclasses was greater than a certain threshold, the pixel was assigned to the peatland superclass. This threshold was set at 25 after conducting 50 experiments and evaluating the resulting peatland boundaries using aerial photographs. For a pixel ultimately to be classified as peatland habitat, an additional condition had to be met: pixels identified as peatland habitat using the spectral threshold approach also had to be part of a cluster of at least 55 contiguous pixels of peatland habitat, which corresponds to a peatland at least 4.95 ha in size. This was done to avoid including many narrow river banks and other small, nonpeatland habitats in the mask. The last step of mask construction consisted of converting groups containing six or fewer contiguous pixels classified as nonpeatland habitat to peatland habitat, providing that they occurred within peatland polygons. The threshold was set at six pixels to exclude mineral outcrops or densely forested peatland islands from our peatland classification.

Distinguishing habitats within peatlands

Once the mask was built, two approaches were tested for classifying within-peatland habitats. The first was a typical multivariate Gaussian-based maximum likelihood (ML) function that assigned to each peatland pixel the most likely habitat class based on the training sites (Richards and Jia 1999). The second, weighted maximum likelihood (WML) approach accounted for positive spatial autocorrelations among neighboring pixels and thereby incorporated the contextual information found in peatland polygons to estimate the most likely class for each pixel. Technically, for every pixel of a given peatland polygon, a set of probabilities that the pixel belonged to each of the habitat classes was first calculated by ML using multinormal functions in which covariance matrices were estimated from the training sites. This information was then used to estimate the relative proportions of each habitat class in the peatland polygons by means of a maximization-expectation algorithm that solved a system of functions adapted from Fortier's (1992) best linear corrector (see Fortier 1999 for details). These proportions or probabilities of occurrence finally served as priors for a Bayesian classification of the pixels pertaining to a given peatland polygon. Therefore, the WML approach was a contextual approach that proceeded on a per-peatland basis, as opposed to the ML approach, which conducted the classification on a per-pixel basis over the entire image in a single step. The two classifications were performed using the CEPIX module of the SURVOL software package (Fortier and Careau 2000). The last image-processing step consisted of rectifying the classified image to a Universal Transverse Mercator map projection using a nearest-neighbor resampling method (Richards and Jia 1999).

Validation of the habitat maps

We validated the results of the two image classifications based on a "blind" ground-truthing that spanned the entire summer of 2000. During that period, we visited 92 peatlands within the study area and nonsystematically recorded 626 DGPS locations using a Trimble Pathfinder Geoexplorer III. We associated a habitat class with each of these locations in the field by visual identification. This was consistent with how we categorized the training sites, that is, qualitatively instead of quantitatively using vegetation surveys.

To assess the performance of our image classifications, we then built an error matrix that compared the habitat classes identified in the field for each of the 626 DGPS locations with those classified on the image using each classification method. We could not calculate classification errors on a pixel-by-pixel basis (Congalton 1991) because (1) according to 40 checkpoints in the field, there were mean (± SD) spatial errors of -1.6 pixels along the east-west axis and 2.0 pixels along the north-south axis on the georeferenced classified image, and (2) the peatland habitats were so narrowly defined that they rarely occurred as large, pure, homogeneous zones. We therefore considered an area of 3 x 3 pixels, for which the bottom-right pixel corresponded to each of the DGPS locations transferred on the georeference-classified image, and contrasted the habitat class identified in the field with the percent cover of each habitat class found within the reference zone of 3 x 3 pixels (0.8 ha). It should be noted that this area was smaller for DGPS locations that occurred near the edges of peatlands. Because we considered nine pixels on the classified image to be comparable to the habitat identification made on a one-pixel basis in the field, we could not calculate commission errors, i.e., verify if the habitat class attributed to each classified pixel corresponded to the habitat class determined in the field. On the other hand, we could still assess the likelihood of omission errors, i.e., the incapacity of the classification method to detect the occurrence of a habitat class in the field.

We assessed the accuracy of our classification method using an additional procedure that compared the accuracy of the results obtained at the 626 DGPS validation locations with that of the results we would have obtained if the reference zones had been randomly distributed among all the peatland polygons. This comparison was intended to show how likely we were to get results as good as the ones we obtained just by chance. More specifically, we calculated how many of the 14,030 random reference zones and how many of the 626 real reference zones had no, one, two, and up to nine pixels of each habitat class. We then built relative-frequency distribution curves that compared the probabilities that each habitat would occur in a given amount in squares measuring 3 x 3 pixels that corresponded to our validation locations and in randomly sampled squares.

Vegetation surveys

Even very accurate maps do not necessarily provide information on how well habitat classes depict the diversity of plant species. For this reason, we carried out detailed vegetation surveys for each of the 13 habitats to determine whether our maps could be regarded as reliable indicators of the occurrence of particular plant species. Vegetation surveys were conducted from 6 June to 7 September, 2000. In total, 252 pixels (mean ± SD = 19.4 ± 0.7 pixels per habitat class) were randomly sampled in 92 different peatlands. For a pixel to be chosen, it had to be surrounded by at least eight pixels of the same habitat class to avoid confounding effects due to spatial error in the field. Site accessibility and field sampling efficiency were additional constraints when selecting the pixels to be sampled, which we located in the field using a DGPS receiver.

For each pixel sampled, we evaluated the vegetation structure in a 20 x 20 m quadrat. Vegetation structure was defined based on 17 strata: water, litter, upright mosses, horizontal mosses, liverworts, lichens, three Sphagnum sections (Acutifolia, Cuspidata, Sphagnum), ericaceous shrubs, Carex, other sedges and forbs, shrubs, birch, pine, spruce, and larch. We estimated the percent cover of these vegetation strata visually while walking across the entire 400-m2 quadrat. Percent cover was assigned to one of seven classes: present, 1–5, 6–10, 11–25, 26–50, 51–75, and 76–100%. The midpoint of each class was used in statistical analyses. We also sampled plant species composition by estimating the percent cover of each species, including mosses, liverworts, and lichens, to the nearest 1% in three circular plots of 0.65 m2 that were nonsystematically distributed within the sampled pixel. We chose these plots by throwing three plastic rings in different directions within the 20 x 20 m sampled area.

Vegetation data analysis

We used a partial ordination approach to assess how well the habitat classes defined a priori for this study reflected the vegetation patterns we observed in the field. More specifically, partial ordinations allowed us to partition the amount of variation in species occurrence and vegetation structure that could be accounted for by the 13 habitat classes defined a priori as well as by certain environmental and spatial variables (Borcard et al. 1992, Legendre and Legendre 1998) such as the depth of the water table, shade cover, tree height, and water chemistry; please see the caption of Fig. 3 for a detailed list. We used Canonical Correspondence Analysis (CCA) and Redundancy Analysis (RDA) to model the relationship between species composition or vegetation structure and the explanatory variables such as habitat classes and environmental and spatial variables, respectively. Ordinations were computed using Canoco 4.0 (ter Braak and Šmilauer 1998). We based our decision about which type of ordination to run on the results of a Detrended Correspondence Analysis (DCA) that showed a large gradient for species data (maximum length of 5.0) and a small one for vegetation structure data (maximum length of 1.8). Ordination methods were thus based on two assumptions for responses to environmental variables and on two corresponding approaches to modeling the species and structural data: a unimodal response with CCA and a linear response with RDA, respectively. We omitted species that had fewer than three occurrences in CCA and log-transformed cover values in CCA and RDA. Environmental variables related to chemical component concentration were log-transformed for both the species (CCA) and the structure (RDA) data analyses. The RDA runs were centered and standardized by species, but not by samples (ter Braak and Šmilauer 1998).

We used the sample scores from the CCA to construct a plot showing the distinctiveness and of the habitat classes based on their species assemblages. We also produced a species bi-plot to identify which environmental variables and individual species were associated with the most distinct habitat classes. For the sample plot and the species bi-plot, we focused the scaling on intersample and interspecies distances, respectively (ter Braak and Šmilauer 1998).


RESULTS

The mask procedure revealed 629 peatland polygons covering a total of 18,103 ha. As seen in Fig. 2, 74% of these polygons were smaller than 20 ha, with 328 of them covering approximately 10 ha and 135 covering between 10 and 20 ha. In addition, 124 (20%) peatlands were between 20 and 100 ha and 27 (4%) between 100 and 200 ha. Only 15 (2%) peatlands were larger than 200 ha, with the largest one covering 756 ha. Fig. 1 shows some examples of peatlands classified using both the maximum likelihood (ML) and the weighted maximum likelihood (WML) procedure. The WML method, which tended to agglomerate isolated pixels, did not produce as much of a "salt and pepper" effect as did the ML classification.

Validation of the habitat maps

We found that the habitat maps produced by the two classification procedures provided a good representation of the the habitats we had identified when we visited the sample of 92 peatlands. The results of the ML method are presented in Table 2, and those of the WML method are given in Table 3. According to the diagonal entries indicating correct classifications, all but one of the habitat classes we identified in the field (represented in the columns) corresponded to the dominant habitats classified in the corresponding 3 x 3 pixel zones (represented in the rows in Table 2). For example, 54% of the 258 pixels corresponding to the 33 locations identified as spruce forest with open canopy (SprFor) in the field were classified as such on the satellite image, a result that is far from negligible when we consider the habitat heterogeneity that characterizes the peatland we surveyed and the extent of the reference zone (Table 2). Lawn with pools (LawnPool) was the only habitat for which the most abundant classified habitat in the reference zones was not of the same type, probably because this habitat was restricted in its distribution and rarely covered areas measuring more than 3 x 3 pixels.

Not only was there good agreement between the habitats identified in the field and the corresponding classified habitats, but also, when there was a lack of fit, it was usually biased toward habitats that are structurally alike. For instance, for the 57 DGPS positions identified in the field as spruce thickets (Spr), the other most abundant habitats classified with the ML procedure in the corresponding 43 ha were herbs with spruce thickets (HerSpr), ericaceous shrubs with spruce thickets (EriSpr), and SprFor; of all the other habitats, these are the most similar to Spr (Table 2).

Independently of the classification procedure (ML or WML), three pairs of habitats seemed to be easily confused with each other: ericaceous shrubs with larch (EriLar) and larch with ericacious shrubs (LarEri), herbs (Herb) and herbs with larch (HerLar), and spruce thickets with pools (SprPool) and SprFor (Tables 2 and 3). We believe that part of this confusion results from the tendency of these habitats to occur together. The two habitats representing larch with herbs (LarHer) and LawnPool are the ones that showed the weakest fit with classified habitats, but again this resulted partly from the fact that these habitats rarely covered an extensive area, so that it was normal to get some contamination within the reference zone of 3 x 3 pixels.

Although the two classification procedures led to similar overall error matrices, there were some differences based on the individual habitat classes. When comparing the numbers on the diagonal between the matrices in Tables 2 and 3, it can be seen that the WML procedure (Table 3) classified some habitats, usually common ones, more accurately than others (Fig. 4). This is because the WML classification procedure takes into account the proportion of each habitat within a peatland polygon when calculating the probability that a pixel will be classified as a certain habitat type. This means that, when a habitat class is rare, the probability that a pixel will be classified as this habitat class is lower than if its relative abundance was not taken into account. However, the WML procedure had the advantage of eliminating the isolated pixels that were common in the classification resulting from the ML procedure (Fig. 1).

The reference zones contained more classified pixels of the habitat class identified in the field than could be expected by chance (Fig. 5). For example, although only 20% of the pixels in the reference zones were classified as LawnPool when this habitat class was identified in the field (Table 2), 50% of these reference zones had between one and six pixels classified as LawnPool habitat. In contrast, it was very unusual to obtain more than one pixel of LawnPool habitat when randomly sampling the peatlands. This situation occurred for all habitat classes.

Both image classifications were probably more accurate than suggested by analysis of the error matrices, because on occasion habitats were misclassified in the field. To assess these potential field identification errors, three persons independently assessed the habitat class at 427 of the 626 validation locations. All three observers disagreed on the classification of 18 (4%) of these locations, whereas two observers agreed on 117 (27%), and all three of them were in agreement on 292 (69%). These agreement rates were better than those that could be expected by chance, with Cohen's kappa indices ranging from 0.61 to 0.79 for pairwise comparisons (Agresti 1996). The assessments by the most experienced fieldworker were used for the analyses. Given this additional information, we are confident that our validation method based on reference zones was appropriate.

Vegetation data analysis

Among the 252 pixels sampled in the field, we recorded a total of 128 species (Appendix 2). When all three groups of explanatory variables, i.e., habitats, environmental variables, and spatial variables, were considered, 31.9% of the variation in species abundance could be explained, compared to 49.9% for the structure data (Fig. 3). This difference stems partly from the fact that there were 93 species but only 17 structure classes. Noise caused by quadrats in which a species was rare or absent was thus more important in the species data than in the structure data. The unbalanced number of response variables between the species and the structure data sets therefore prevented the direct comparison of the amount of variation explained by the habitat classes for these two data sets. We accordingly had to quantify the explanatory power of the habitat classes on a relative basis with the environmental and spatial variables.

Variance partitioning via canonical correspondence analysis (CCA) revealed that, of the 31.9% of the variation in the species data explained by the three groups of variables, 9.6% could be attributed solely to habitat classes, 9.5% to environmental variables, and only 0.9% to spatial variables; the three groups shared 1.5% of the variation (Fig. 3). For the structure data analysis, redundancy analysis (RDA) showed that, of the 49.9% of the total variation explained by the three groups of variables, up to 14.3% could be attributed solely to habitat classes, compared to 8.7% for environmental variables and only 1% for spatial variables; the three groups shared 3.1% of the variation (Fig. 3). These results indicate that the 13 habitat classes defined a priori were as good as the 15 environmental variables at explaining species distribution, whereas, for vegetation structure patterns, the habitat classes performed 1.6 times better than did the environmental variables.

According to the CCA, the six habitats dominated by ericaceous shrubs or forested habitats were similar in terms of their species assemblages (Graph A in Fig. 6). For example, species common to all these groups were Kalmia angustifolia, Ledum groenlandicum, Polytrichum strictum, Sphagnum fuscum, and S. capillifolium (Graph B in Fig. 6, Appendix 2). Habitats with pools (SprPool, LawnPool) and associated species Drosera rotundifolia, Nuphar lutea ssp. variegata, Rhynchospora alba, Utricularia cornuta, Cladopodiella fluitans, and Sphagnum cuspidatum were well-separated from the previous group. Although there was less contrast than in habitats with pools, habitats with herbs also presented quite distinct species assemblages, especially the habitats designated as Herb, tall sedges, and LarHer (Graph A in Fig. 6, Appendix 2). These herb-dominated habitats tended to be richer in mineral elements such as Ca++, Mg++, and Fe++, less acidic, and wetter than other habitats, which favored vascular species such as Carex exilis, C. limosa, Menyanthes trifoliata, Pogonia ophioglossoides, and Scheuchzeria palustris as well as Sphagnum species including S. fallax, S. majus, and S. papillosum (Graph B in Fig. 6; Appendices 1, 2, and 3). Herb-dominated habitats were also the most variable, as shown by the error bars in Graph A of Fig. 6. In contrast, the species composition of habitats dominated by ericaceous shrubs was more likely to be predicted. Even though the habitat LarEri was not highly distinct from the rest of the habitats, it tended to be located near the edges of the peatlands. Consequently, this type of habitat was richer in base cations as measured by higher concentrations of Ca++, and it sheltered species such as Nemopanthus mucronatus and Viburnum nudum var. cassinoides (Graph B in Fig. 6, Appendices 2 and 3).


DISCUSSION

To our knowledge, this study is one of the first to produce a high-resolution map of peatland habitats using satellite imagery. Our results show that it is possible to map as many as 13 peatland habitats with Landsat 7/ETM+ data and to represent peatland vegetation with a level of accuracy that is probably compatible with regional investigation and conservation objectives. Moreover, our classification methodology, which included the use of a masking procedure to isolate peatland polygons from the rest of the image before classifying within-peatland habitats, made it possible to characterize small peatlands (< 40 ha) that would have been ignored by the traditional field surveys used to produce peatland atlases (e.g., Buteau 1989). However, our study also indicates a need for caution when associating peatland habitats with polygons smaller than 10 ha, because some land-cover types can produce a similar radiance and thus introduce noise into the mapping process. For instance, we noted that we had mistakenly classified as peatland habitat certain small polygons located under hydroelectric lines or on river banks that had only a few of the characteristics of peatland, e.g., moist ground with several patches of sphagnum moss. For this reason, more detailed ground-truthing should be conducted before small peatland polygons can be considered as potential conservation units. Nevertheless, small polygons should not be ignored for peatland conservation purposes, because small peatlands can have a great diversity of plant species; kettle holes (e.g., Lindholm and Vasander 1983) are a good example of this.

Maximum likelihood vs. weighted maximum likelihood:
comparing the classification procedures

For our purposes, the weighted maximum likelihood (WML) procedure, a contextual classifier that takes into account the radiance value of the surrounding pixels, was, on the whole, no more accurate than the standard maximum likelihood (ML) approach; both classification methods had approximately the same overall level of accuracy. Similarly, studies that used a contextual classifier (e.g., Thunnissen et al. 1992, Hubert-Moy et al. 2001) generally improved their accuracy by only a few percentage points. However, in our case, common habitats were more accurately classified using the WML procedure, whereas rare habitats were better represented using the ML procedure. This was partly because the WML procedure led to a more homogeneous classification that contained fewer isolated pixels or small clusters of pixels of any given habitat class. Because of the spatial error in the georeferenced classification, we could not assess directly through our validation procedure whether the loss of these individual pixels constituted an improvement or not. However, based on our field experience, we believe that some habitats are unlikely to occur on an individual-pixel basis within a contrasting habitat class, especially if their hydrological states are different. Consequently, even though the WML procedure underestimates the rarest habitats, it avoids mapping less probable habitats and eliminates salt-and-pepper noise. When it comes to the identification of potential conservation units, the WML classification procedure is preferable, because it avoids classifying rare habitats whose spectral distinctiveness is not high enough to compensate for their small extent. Given that most conservation approaches aim to maximize the complementarity of sites (Pressey et al. 1993), the site-selection algorithms would then have to search for the sites in which rare habitats are more extensive, ensuring that the whole range of habitats is more likely to be preserved.

Validation constraints

The combined problems of the high heterogeneity of peatland habitats at small spatial scales (< 1 ha) and of spatial error on the classified georeferenced image prevented us from conducting the accuracy assessment (i.e., building an error matrix) on a pixel-by-pixel basis. Aaviksoo et al. (2000), who faced the same problems when classifying 11 peatland habitats in Estonia, opted not to evaluate systematically the accuracy of their classification and gave only approximate estimations. We first tried to circumvent the problem of location error by placing the field validation points in relatively homogeneous zones, and this technique, even though it limits the assessment to homogeneous zones, can lead to positively biased results (Hammond and Verbyla 1996). However, we did not succeed using this strategy, because most of our validation points were in heterogeneous zones containing at least two classes of habitats. In fact, our random sampling of 14,030 squares of 3 x 3 pixels on the image classified using the ML procedure and corresponding to 0.8 ha in the field showed that only a small proportion (13%) of these reference zones contained only one habitat class. To our knowledge, very few published studies have conducted classification work with such a large number of heterogeneous classification units.

Studies on classification accuracy assessment have acknowledged that spatial errors can lead to an underestimation of the spectral accuracy of classified images, particularly in the presence of a large number of heterogeneous classification units (Hammond and Verbyla 1996, Stehman and Czaplewski 1998). Despite this, only cluster sampling has so far been used to compensate for the impossibility of assessing the accuracy of a classification on a pixel-by-pixel basis (Cibula and Nyquist 1987, Stenback and Congalton 1990, Watson and Wilcock 2001). This method is, in fact, pseudo-cluster sampling and not true cluster sampling, which is used to reduce the costs of evaluating map accuracy (Todd et al. 1980, Martin 1989, Martin and Howarth 1989, Stehman and Czaplewski 1998). It consists of sampling clusters of 3 x 3 or 5 x 5 classified pixels and considers the latter properly classified when the dominant class of the clusters, or at least one pixel of the clusters, corresponds to the class observed in the field. This "dichotomized" categorization of clusters, however, does not take advantage of the information conveyed by the specific composition of the clusters (i.e., the amount of cover of each class in a cluster). Again, in our efforts to circumvent the spatial error problem, we applied a method that compared the likelihood of observing a certain amount of a given habitat class in the clusters used to validate the classifications and in clusters selected at random from all the peatland polygons. These comparisons indicated that the validation clusters contained more of the targeted habitat than did randomly selected clusters; this was the case for all of our habitat classes, which increased our confidence that our error matrices were meaningful. This additional analysis conducted to complement our nonstandard error matrices can be seen as one answer to Congalton's (1996: 127) plea that " ... work is needed to go beyond the error matrix and introduce techniques that build upon the information in the matrix and make it more meaningful ... "

Vegetation patterns

Because ecological mapping procedures based on supervised classification involve the definition of habitat classes a priori, they are inevitably somewhat subjective. Even though supervised classification techniques are widely applied in remote sensing studies, the explanatory power of the different habitat classes with respect to relevant ecological attributes such as species composition or vegetation structure is rarely quantified. We nevertheless are of the opinion that such an evaluation is critical if the results of the classification are going to be applied to land management or conservation problems. The detailed vegetation surveys we conducted enabled us to proceed with this type of assessment, which revealed that the 13 habitat classes were defined precisely enough to explain a large part of the variation in plant species assemblages and vegetation structure at least as effectively as this variation could be accounted for by the 15 measured environmental variables. We therefore believe that our habitats are good surrogates for peatland management and conservation purposes, at least with regard to vegetation diversity.

In addition, the detailed vegetation surveys allowed us to determine the range of variation among habitats in terms of plant species assemblages and vegetation structure. This information is of considerable value when it comes to deciding whether all habitats deserve an equal level of protection and determining the amount of habitat to set aside for conservation. In fact, information about the distinctiveness and variability of habitats can be coupled with other criteria, such as habitat rarity and vulnerability, to derive some habitat-specific weighting factor that can be used to set conservation targets, such as the proportion of initial habitat cover to be protected (see Pressey and Taffs 2001). For example, peatland habitats with pools might be favored if conservation procedures were to be be implemented in our study area because of their high distinctiveness and low aerial extent (Poulin et al. 1999). Moreover, preference might also be given to habitats dominated by herbs, especially to the classes for herbs (Herb), tall sedges (TSedge), and larch with herbs (LarHer), because they are relatively distinct and present more variation in their species composition. This variation might complement species diversity as a measure to be taken into account for conservation purposes, because these two variables are not necessarily correlated. Indeed, Vitt et al. (1995) found that, although fens were not individually richer in bryophyte species than bogs, there was more variability from site to site, leading to a higher diversity of species when a group of fen sites is considered. In our case, more peatlands containing herb habitats should be preserved if we want to capture the different species associated with their high β diversity (Whittaker 1977, Noss 1983). Consequently, detailed surveys not only help to assess classification accuracy but also provide an additional tool when setting conservation priorities to maximize biodiversity in selected reserves. Because peatlands are dynamic systems evolving in time from fens to bogs, successional sequences should also be considered when planning conservation actions.


CONCLUSION

The final product of classification studies like this one is a map of landscape features, in our case 13 peatland habitats classified on the basis of Landsat 7/ETM+ data. Although we obtained quite similar results whether we used conventional or weighted classification procedures based on maximum likelihood (ML) functions, the weighted maximum likelihood (WML) procedure was found to classify common habitats more accurately than the ML procedure, and the converse was true for rare habitats. Nevertheless, we believe that the WML procedure is more reliable when it comes to the identification of potential conservation units, because it provides a more robust identification of rare habitats and thus minimizes the chance of misallocating protected sites. Moreover, as sensor spatial resolution improves, contextual classifiers will become standard procedures to avoid an increase in the salt-and-pepper effect.

Another important issue that affects the meaningful application of the results of ecological maps is that habitat classes reflect ecological attributes that are relevant to the land management and conservation problems at hand. Along these lines, our ordination approach revealed that the peatland habitats defined a priori for the supervised classification procedures were representative of the species distribution and vegetation structure patterns in the peatlands of the mid-St. Lawrence plain of southern Quebec. Finally, and perhaps more importantly, our results are readily accessible to land planners and conservationists seeking to establish nature reserve networks. Indeed, many of the site-selection algorithms developed during the past two decades require this type of spatial database (Pressey et al. 1995, Pressey and Taffs 2001). Furthermore, this need is likely to increase because new site-selection algorithms that take into account the persistence of populations are currently being developed and will ultimately require spatially explicit data (Cabeza and Moilanen 2001).


RESPONSES TO THIS ARTICLE

Responses to this article are invited. If accepted for publication, your response will be hyperlinked to the article. To submit a comment, follow this link. To read comments already accepted, follow this link.


Acknowledgments:

Part of the work for this paper was done while the corresponding author was a member of the faculty of the Department of Forest Ecology, University of Helsinki, P.O. Box 27, Helsinki 00014, Finland (phone: +358-50-365-2232, fax: +358-9-191-58100), and she would like to thank the University of Helsinki for its support. The authors are also grateful to V.-A. Anger, R. Cliche, D. Cobbaert, B. Drolet, M. Girard, O. Soucy, and R. Rowland for field assistance; D. Bastien for his help with lichen, liverwort, and moss identifications; C. Boismenu for revising Appendix 2; R. Gauthier for his contribution with vascular plant identification; Dr. C. Cassady St. Clair for her help with logistics; and the Ministère des Ressources naturelles du Québec (MRNQ) for providing aerial photographs. We are also indebted to D. Rochefort from the MRNQ who georefenced the classified image and helped, with C. Seuthé, to purchase the satellite image. We thank J.-J. Fortier for his explanations of statistical estimations and classification techniques. Dr. M. Bélisle, S. Jauhiainen, and Dr. J. Päivänen as well as two anonymous reviewers provided helpful comments on earlier versions of the manuscript. This research has been supported by the Ministère de l'Environnement du Québec (programme PARDE), the Province of Quebec Society for the Protection of Birds, and the Society of Wetland Scientists. MP benefited from NSERC (Canada) and FCAR (Quebec) scholarships.


LITERATURE CITED

Aaviksoo, K., J. Paal, and T. Dislis. 2000. Mapping of wetland habitat diversity using satellite data and GIS: example from the Alam-Pedja nature reserve, Estonia. Proceedings of the Estonian Academy of Sciences: Biology, Ecology 49:177-193.

Agresti, A. 1996. An introduction to categorical data analysis. John Wiley, New York, New York, USA.

Anderson, L. E. 1990. A checklist of Sphagnum in North America north of Mexico. The Bryologist 93:500-501.

Anderson, L. E., H. A. Crum, and W. R. Buck. 1990. List of the mosses of North America north of Mexico. The Bryologist 93:448-499.

Boone, R. B., and W. B. Krohn. 1999. Modeling the occurrence of bird species: are the errors predictable? Ecological Applications 9:835-848.

Borcard, D., P. Legendre, and P. Drapeau. 1992. Partialling out the spatial component of ecological variation. Ecology 73:1045-1055.

Boresjö Bronge, L., and B. Nöslund-Landenmark. 1999. Satellite data for wetland classification: spectral characteristics of Swedish mire vegetation. Report for the National Land Survey and the Swedish National Space Board, Swedish Space Corporation, Number X-Publ-41. Swedish Space Corporation, Solna, Sweden.

Brummit, R. K., and C. E. Powell, editors. 1992. Authors of plant names. Royal Botanic Gardens, Kew, London, UK.

Budd, J. T. C., and E. J. Milton. 1982. Remote sensing of salt marsh vegetation in the first four proposed Thematic Mapper bands. International Journal of Remote Sensing 3:147-161.

Buteau, P. 1989. Atlas des tourbières du Québec méridional. Gouvernement du Québec, Quebec City, Quebec, Canada.

Cabeza, M., and A. Moilanen. 2001. Design of reserve networks and the persistence of biodiversity. Trends in Ecology and Evolution 16:242-248.

Cibula, W. G., and M. O. Nyquist. 1987. Use of topographic and climatological models in a geographical data base to improve Landsat MSS classification for Olympic National Park. Photogrammetric Engineering and Remote Sensing 53:67-75.

Congalton, R. G. 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of the Environment 37:35-46.

Congalton, R. G. 1996. Accuracy assessment: a critical component of land cover mapping. Pages 119-131 in J. M. Scott, T. H. Tear, and F. W. Davis, editors. Gap analysis: a landscape approach to biodiversity planning; proceedings of the ASPRS/GAP Symposium. American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland, USA.

Debinski, D. M., K. Kindscher, and M. F. Jakubauskas. 1999. A remote sensing and GIS-based model of habitats and biodiversity in the Greater Yellowstone Ecosystem. International Journal of Remote Sensing 20:3281-3291.

Environmental Systems Research Institute. 1996. ArcView GIS V. 3.1.1. ESRI, Redlands, California, USA.

Esslinger, T. L., and R. S. Egan. 1995. A sixth checklist of the lichen-forming, lichenolous, and allied fungi of the continental United States and Canada. The Bryologist 98:467-549.

Fortier, J.-J. 1992. Best linear corrector of classification estimates of proportions of objects in several unknown classes. Canadian Journal of Statistics 20:23-33.

Fortier, J.-J. 1999. Une approche statistique pratique pour le traitement de la non-réponse dans les sondages: l'estimation de vraisemblance maximum privée. 67ième Congrès de l'Association canadienne francaise pour l'avancement des sciences (Ottawa, 1999). Association canadienne francaise pour l'avancement des sciences, Montreal, Quebec, Canada. Available online at http://www.acfas.ca/congres/congres67/C2105.HTM.

Fortier, J.-J., and D. Careau. 2000. SURVOL v.2000. Société de Mathématiques Apppliquées, Quebec City, Quebec, Canada.

Franklin, J. F. 1993. Preserving biodiversity: species, ecosystems, or landscapes? Ecological Applications 3:202-205.

Fuller, R. M., G. B. Groom, S. Mugisha, P. Ipulet, D. Pomeroy, A. Katende, R. Bailey, and R. Ogutu-Ohwayo. 1998. The integration of field survey and remote sensing for biodiversity assessment: a case study in the tropical forests and wetlands of Sango Bay, Uganda. Biological Conservation 86:379-391.

Fuller, R. M., R. J. Parsell, M. Oliver, and G. Wyatt. 1989. Visual and computer classifications of remotely-sensed images; a case study of grasslands in Cambridgeshire. International Journal of Remote Sensing 10:193-210.

Glaser, P. H. 1983. Vegetation patterns in the North Black River peatland, northern Minnesota. Canadian Journal of Botany 61:2085-2104.

Glaser, P. H. 1989. Detecting biotic and hydrogeochemical processes in large peat basins with Landsat TM imagery. Remote Sensing of Environment 28:109-119.

Glaser, P. H., G. A. Wheeler, E. Gorham, and H. E. Wright, Jr. 1981. The patterned mires of the Red Lake peatland, northern Minnesota: vegetation, water chemistry and landforms. Journal of Ecology 69:575-599.

Gorham, E. 1990. Biotic impoverishment in northern peatlands. Pages 65-98 in G. M. Woodwell, editor. The earth in transition: patterns and processes of biotic impoverishment. Cambridge University Press, Cambridge, UK.

Grenier, M., D. Bordage, and N. Plante. 1994. La télédétection, un complément avantageux aux inventaires pour évaluer la répartition de la sauvagine sur de vastes territoires. Canadian Journal of Remote Sensing 20:162-170.

Hammond, T. O., and D. L. Verbyla. 1996. Optimistic bias in classification accuracy assessment. International Journal of Remote Sensing 17:1261-1266.

Herr, A. M., and L. P. Queen. 1993. Crane habitat evaluation using GIS and remote sensing. Photogrammetric Engineering and Remote Sensing 59:1531-1538.

Hogson, M. E., J. R. Jensen, H. E. Mackey, and M. C. Coulter. 1987. Remote sensing of wetland habitat: a wood stork example. Photogrammetric Engineering and Remote Sensing 53:1075-1080.

Holopainen, M., and G. Wang. 1998. The calibration of digitized aerial photographs for forest stratification. International Journal of Remote Sensing 19:677-696.

Hubert-Moy, L., A. Cotonnec, L. Le Du, A. Chardin, and P. Perez. 2001. A comparison of parametric classification procedures of remotely sensed data applied on different landscape units. Remote Sensing of Environment 75:174-187.

Hughes, J. B., G. C. Daily, and P. R. Ehrlich. 2000. Conservation of insect diversity: a habitat approach. Conservation Biology 14:1788-1797.

Joint, I., and S. B. Groom. 2000. Estimation of phytoplankton production from space: current status and future potential of satellite remote sensing. Journal of Experimental Marine Biology 250:233-255.

Jones, B., and D. LeAnn. 2000. Landsat 7 processing on the EROS Data Center's National Land Archive Processing System (NLAPS) in algorithms for multispectral, hyperspectral, and ultraspectral imagery VI. Pages 319-327 in S. S. Shen and M. R. Descour, editors. Proceedings of the Society of Photo-Optical Instrumentation Engineers, Volume 4049. Society of Photo-Optical Instrumentation Engineers, Bellingham, Washington, USA.

Kartesz, J. T. 1994. A synonymized checklist of the vascular flora of the United States, Canada, and Greenland. Volume 2. Second edition. Timber Press, Portland, Oregon, USA.

Kasischke, E. S., J. M. Melack, and M. C. Dobson. 1997. The use of imaging radars for ecological applications—a review. Remote Sensing of Environment 59:141-156.

Konrad, S.-L. R., and R. S. Rempel. 1990. Cost-effectiveness of Landsat TM classification by operations staff. IEEE Transactions on Geoscience and Remote Sensing 28:769-771.

Lappalainen, E., editor. 1996. Global peat resources. International Peat Society, Jyväskylä, Finland.

Lauver, C. L. 1997. Mapping species diversity patterns in the Kansas shortgrass region by integrating remote sensing and vegetation analysis. Journal of Vegetation Science 8:387-394.

Lauver, C. L., and J. L. Whistler. 1993. A hierarchical classification of Landsat TM imagery to identify natural grassland areas and rare species habitat. Photogrammetric Engineering and Remote Sensing 59:627-634.

Legendre, P., and L. Legendre. 1998. Numerical ecology. Second English edition. Elsevier Science, New York, New York, USA.

Lewis, M. M. 1998. Numeric classification as an aid to spectral mapping of vegetation communities. Plant Ecology 136:133-149.

Lindholm, T., and H. Vasander. 1983. Bryophyte flora in an esker kettle hole mire, Heinisuo, southern Finland. Memoranda Societatis pro Fauna et Flora Fennica 59:144-150.

Margules, C. R., and R. L. Pressey. 2000. Systematic conservation planning. Nature 405:243-253.

Martin, L. R. G. 1989. Accuracy assessment of Landsat-based visual change detection methods applied to the rural-urban fringe. Photogrammetric Engineering and Remote Sensing 55:209-215.

Martin, L. R. G., and P. J. Howarth. 1989. Change-detection accuracy assessment using SPOT multispectral imagery of the rural-urban fringe. Remote Sensing of Environment 30:55-66.

Mumby, P. J., E. P Green, A. J. Edwards, and C. D. Clark. 1999. The cost-effectiveness of remote sensing for tropical coastal resources assessment and management. Journal of Environmental Management 55:157-166.

Nagendra, H., and M. Gadgil. 1999. Satellite imagery as a tool for monitoring species diversity: an assessment. Journal of Applied Ecology 36:388-397.

Nøhr, H., and A. F. Jørgensen. 1997. Mapping of biological diversity in Sahel by means of satellite image analyses and ornithological surveys. Biodiversity and Conservation 6:545-566.

Noss, R. F. 1983. A regional landscape approach to maintain diversity. BioScience 33:700-706.

Noss, R. F. 1987. From plant communities to landscapes in conservation inventories: a look at the nature conservancy (USA). Biological Conservation 41: 11-37.

Noss, R. F. 1996. Ecosystems as conservation targets. Trends in Ecology and Evolution 11:351.

Oetter, D. R., W. B. Cohen, M. Berterretche, T. K. Maiersperger, and R. E. Kennedy. 2001. Land cover mapping in an agricultural setting using multiseasonal Thematic Mapper data. Remote Sensing of Environment 76:139-155.

Palylyk, C. L., P. H. Crown, and L. W. Turchenek. 1987. Landsat MSS data for peatland inventory in Alberta. Pages 365-371 in C. D. A. Rubec and R. P. Overend, editors. Symposium '87, wetlands/peatlands. Canadian Society of Wetlands and Peatlands and Canadian National Committee, International Peat Society, Edmonton, Alberta, Canada.

Panzer, R., and M. W. Schwartz. 1998. Effectiveness of a vegetation-based approach to insect conservation. Conservation Biology 12:693-702.

Pearce, J., and S. Ferrier. 2001. The practical value of modelling relative abundance of species for regional conservation planning: a case study. Biological Conservation 98:33-43.

Peñuelas, J., J. A. Gamon, K. L. Griffin, and C. B. Field. 1993. Assessing community type, plant biomass, pigment composition, and photosynthetic efficiency of aquatic vegetation from spectral reflectance. Remote Sensing of Environment 46:110-118.

Poulin, M., L. Rochefort ,and A. Desrochers. 1999. Conservation of bog plant species assemblages: assessing the role of natural remnants in mined sites. Applied Vegetation Science 2:169-180.

Pressey, R. L. 1994. Land classifications are necessary for conservation planning but what do they tell us about fauna? Pages 31-41 in D. Lunney, S. Hand, P. Reed, and D. Butcher, editors. Land classifications are necessary for conservation planning but what do they tell us about fauna? Royal Zoological Society of New South Wales, Mosman, Australia.

Pressey, R. L., S. Ferrier, C. D. Hutchinson, D. P. Sivertsen, and G. Manion. 1995. Planning for negotiation: using an interactive geographic information system to explore alternative protected area networks. Pages 23-33 in D. A. Saunders, J. L. Craig, and E. M. Mattiske, editors. Nature conservation: the role of networks. Surrey Beatty, Sydney, Australia.

Pressey, R. L., C. J. Humphries, C. R. Margules, R. I. Vane-Wright, and P. H. Williams. 1993. Beyond opportunism: key principles for systematic reserve selection. Trends in Ecology and Evolution 8:124-128.

Pressey, R. L., and K. H. Taffs. 2001. Scheduling conservation action in production landscapes: priority areas in western New South Wales defined by irreplaceability and vulnerability to vegetation loss. Biological Conservation 100:355-376.

Price, J. 1997. Soil moisture, water tension, and water table relationships in a managed cutover bog. Journal of Hydrology 202:21-32.

Quarmby, N. A., F. Everingham, and E. Reid. 1997. Scottish blanket bog inventory: Lewis and Harris—characterisation of blanket bogs using Landsat Thermatic Mapper. Scottish Natural Heritage Research, Survery and Monitoring Report 85. Scottish Natural Heritage, Edinburgh, UK.

Rey-Benayas, J. M., and K. O. Pope. 1995. Landscape ecology and diversity patterns in the seasonal tropics from Landsat TM imagery. Ecological Applications 5:386-394.

Richards, J. A., and X. Jia. 1999. Remote sensing digital image analysis: an introduction. Springer-Verlag, Berlin, Germany.

Robitaille, A., and J.-P. Saucier. 1998. Paysages régionaux du Québec méridional. Gouvernement du Québec, Quebec City, Quebec, Canada.

Roughgarden, J., S. W. Running, and P. A. Matson. 1991. What does remote sensing do for ecology? Ecology 72:1918-1922.

Schoch, G. C., and M. N. Dethier. 1996. Scaling up: the statistical linkage between organismal abundance and geomorphology on rocky intertidal shorelines. Journal of Experimental Marine Biology and Ecology 201:37-72.

Schwartz, M. W. 1999. Choosing the appropriate scale of reserves for conservation. Annual Review of Ecology and Systematics 30:83-108.

Scott, J. M., F. Davis, B. Csuti, R. Noss, B. Butterfield, C. Groves, H. Anderson, S. Caicco, F. D'Erchia, T. C. Edwards Jr., J. Ulliman, and R. G. Wright. 1993. Gap analysis: a geographic approach to protection of biological diversity. Wildlife Monographs 123:1-41.

Sjørs, H. 1952. On the relation between vegetation and electrolytes in north Swedish mire waters. Oikos 2:241-258.

Šmilauer, P. 1992. CanoDraw 3.0 users' guide. Version 3.0. Microcomputer Power, Ithaca, New York, USA.

Spanglet, H. J., S. L. Ustin, and E. Rejmankova. 1998. Spectral reflectance characteristics of California subalpine marsh plant communities. Wetlands 18:307-319.

Stehman, S. V., and R. L. Czaplewski. 1998. Design and analysis for thematic map accuracy assessment: fundamental principles. Remote Sensing of Environment 64:331-344.

Stenback, J. M., and R. G. Congalton. 1990. Using Thematic Mapper imagery to examine forest understory. Photogrammetric Engineering and Remote Sensing 56:1285-1290.

Stoms, D. M., and J. E. Estes. 1993. A remote sensing research agenda for mapping and monitoring biodiversity. International Journal of Remote Sensing 14:1839-1860.

Stotler, R., and B. Crandall-Stotler. 1977. A checklist of the liverworts and hornworts of North America. The Bryologist 80:405-428.

Tatu, K., M. M. Kimothi, and J. S. Parihar. 1998. Methodology for remote sensing of waterbird habitats in an inland wetland. Indian Forester 124:841-852.

ter Braak, C. J. F., and P. Šmilauer. 1998. Canoco release 4: reference manual and user's guide to Canoco for Windows: software for Canonical Community Ordination. Microcomputer Power, Ithaca, New York, USA.

Thunnissen, H. A. M., M. N. Jaarsma, and O. F. Schoumans. 1992. Land cover inventory in the Netherlands using remote sensing; application in a soil and groundwater vulnerablitiy assessment system. International Journal of Remote Sensing 13:1693-1708.

Todd, W. J., D. G. Gehring, and J. F. Haman. 1980. Landsat wildland mapping accuracy. Photogrammetric Engineering and Remote Sensing 46:509-520.

Van Seters, T. E., and J. S. Price. 2001. The impact of peat harvesting and natural regeneration on the water balance of an abandoned cutover bog, Quebec. Hydrological Processes 15:233-248.

Verry, E. S. 1997. Hydrological processes of natural, northern forested wetlands. Pages 61-70 in C. C. Trettin, M. F. Jurgensen, D. F. Grigal, M. R. Gale, and J. K. Jeglum, editors. Northern forested wetlands: ecology and management. Lewis, Boca Raton, Florida, USA.

Vitt, D. H., Y. Li, and R. J. Belland. 1995. Patterns of bryophyte diversity in peatlands of continental western Canada. The Bryologist 98:218-227.

Watson, N., and D. Wilcock. 2001. Preclassification as an aid to the improvement of thematic and spatial accuracy in land cover maps derived from satellite imagery. Remote Sensing of Environment 75:267-278.

Whittaker, R. H. 1977. Evolution of species diversity in land communities. Evolutionary Biology 10:1-67.

Wilcove, D. 1993. Getting ahead of the extinction curve. Ecological Applications 3:218-220.

Zhang, M., S. L. Ustin, E. Rejmankova, and E. W. Sanderson. 1997. Monitoring Pacific Coast salt marshes using remote sensing. Ecological Applications 7:1039-1053.


Address of Correspondent:
Monique Poulin
Groupe de recherche en écologie des tourbières
Centre d'études nordiques
Faculté des sciences de l'agriculture et de
l'alimentation
Pavillon Paul-Comtois
Université Laval
Quebec City, Quebec, Canada G1K 7P4
Phone: +359-9-58115
Fax: +358-9-191-58100



Home | Archives | About | Login | Submissions | Notify | Contact | Search