this study euclidian (pythagorean) distance was determined to be appropriate since it ... heterogeneity within groups equals expectation by chance then A = 0.
Ecological classifications derived from spectral and vegetation data for Cape Bounty, Melville Island David M. Atkinson and Paul Treitz Department of Geography Queen’s University, Kingston, Ontario, Canada, K7L 3N6
ABSTRACT Vegetation is both an integrator and indicator of climate and ecosystem properties. Discerning the pattern of vegetation can provide a connection to the patterns of carbon flux. It may be possible to measure ecosystem processes in common vegetation communities, at the plot level, and extrapolate them over a larger area using spatially-continuous remote sensing data. In the arctic environment where vegetation is highly spatially variable, the use of high resolution imagery can help in discerning the patterns of vegetation and ecosystem processes. The primary objective of this research is to explore a link between the theories and practices of classification of vegetation data by ecologists and image classification for mapping vegetation by remote sensing scientists. This study looks to develop a methodology of relating ecological ordination and classifications techniques, derived using species and cover abundance data, along with measured environmental variables, from Cape Bounty, Melville Island, Nunavut, with remotely-sensed data. Ordination techniques are used to determine the natural arrangement of sample sites followed by cluster analysis to create ecological classes. Multi-response permutation procedure (MRPP) is applied to compare clusters. The derived cluster classes are then used to classify high spatial resolution IKONOS imagery. Ordination, clustering, and classification results showed moderate levels of success. Correspondence analysis (CA) cluster classifications performed slightly better (overall accuracy = 70.9%) than CCA classifications (overall accuracy = 66.2%). The results of this study illustrate that combination of ecological and remote sensing techniques can produce classifications that are ecologically meaningful and spectrally significant in the arctic environment. Keywords: Arctic, Vegetation Mapping, Ordination, Clustering, Remote Sensing
1
Introduction
Changes to the Arctic climate have been observed over the past century (Serreze et al. 2000). These changes could be manifested through shifts in vegetation phenology and species composition. Arctic warming is expected to: (1) promote plant growth and sequestration of carbon from the atmosphere; and (2) increase soil microbial respiration rates, releasing more carbon to the atmosphere (Stieglitz et al. 2000). If ecosystem respiration (ER) exceeds gross ecosystem production (GEP) a positive feedback loop could occur, thus intensifying global climate change.
Within the arctic ecosystem, processes and community types are highly spatially variable. Detailed community level knowledge along with high resolution remote sensing can provide us with the ability to understand fine-grain spatial variation and improve our ability to scale to synoptic predictions. Knowledge obtained through detailed studies at local sites can be used to develop inputs to models of arctic ecosystem processes from community to landscape scales. Specifically, remote sensing has the potential to provide valuable information for the assessment and monitoring of vegetation patterns which can be utilized to predict patterns of carbon flux. Currently, there exists a need to establish and clarify the link between theories and practices of classification by ecologists and remote sensing scientists (Thomas et al. 2002). Vegetation analysis is designed to highlight relationships between stands, species, or both (Causton 1988). Plant ecologist often apply a methodological duet of cluster analysis and ordination to organize species data into discrete ‘‘associations’’ that offer valuable information about the relationships among species and the ecological processes occurring within a community. The decision to use one method over the other has often posed a philosophical dilemma of whether vegetation is organized as a continuum or as discrete clusters. Each method has its problems and limitations, yet they may be viewed as complementary, and when applied together, offer useful information about species relationships and distributions (Thomas et al. 2002). Remote sensing has the ability to provide spatially-continuous data regarding vegetation and terrain patterns, in a range of spatial, spectral, and temporal resolutions, which can be used for observing, investigating, and analyzing biophysical properties of vegetation at various landscape scales (Tieszen et al. 1997; Stow et al. 1998, Stow et al. 2000). Biophysical remote sensing and the characterization of the spatial distribution of ecological classes are based on the assumption of unique spectral characteristics of species and species associations. The goal of such a link between ecologists and remote sensing scientists would be to classify vegetation communities into statistically derived, spectrally significant, ecologically meaningful units. A common derivative of optical, multispectral, remotely sensed data is a classified map of vegetation communities (Stow 1998). Image classification techniques are well documented in the literature and there have been many attempts at using remotely sensed data to produce vegetation maps of arctic vegetation communities (Stow et al. 1993; Shippert et al. 1995; Gould 2000, Walker et al. 2005). The emphasis of remote sensing research for ecological purposes has long been on vegetation structure, cover, and temporal dynamics, attributes that are clearly affect spectral reflectance in the visible and near infrared wavelengths (Lewis 1998). Surprisingly little attention has been given to the nature of relationships between spectral classes and vegetation classes that are of interest to the ecologist, a discontinuity which has been remarked by several authors (Graetz 1989; Roughgarden et al. 1991; Wickland 1991). In remote sensing classifications these relationships are often not made explicit, and more attention is given to spectral characteristics of the mapping units; often at the expense of the description and quantification of their ecological characteristics (Lewis, 1998). There have been some notable attempts to demonstrate the nature of relationships between spectral and vegetation classes with varying techniques and levels of success. Approaches have included, the use of multiple discriminant analysis to predict community type on the basis of spectral variables (Ustin et al. 1986., Pando et al. 1992; Hobbs et al. 1989; Ghitter et al. 1995), statistical association between the distributions of spectral classes and independently mapped vegetation classes (Price et al. 1992), and statistical comparisons between spectral and either numeric or subjective vegetation classifications (Catt et al. 1987; Toth et al. 1991; Lewis 1994). Lewis’s (1998) approach had these key elements: (a) the use of cover rather than density or presence/absence to quantify the vegetation, (b) the inclusion of physical components as well as vegetation cover to describe and classify field sites, (c) development of an objective land cover classification from this quantitative data, (d) use of the field sample sites as training areas for the spectral classification, and (e) the use of a discriminant function to effectively tie the two classifications together. One of the important elements is the inclusion of physical as well as biotic ground cover components to characterise the vegetation. Lewis (1998) states that it is particularly important in the study sites of arid environments since
perennial and ephemeral vegetation rarely cover over 30%, and the underlying soils and rocks form a significant portion of the landscape. This may hold true for other sparsely vegetated areas, such as the high Arctic. The use of soil and rock, as well as vegetation variables, to classify the landscape allows segregation of sites that may have the same dominants, but which vary in relative cover, a difference also likely to be captured by image reflectance, and which is relevant for management of the land (Lewis, 1998). The goal of this study is to explore a methodology for the creation an ecologically meaningful vegetation classification of high spatial resolution remotely-sensed data for Cape Bounty, Melville Island, Nunavut, Canada that could be used in the generation of model estimates of trace gas flux for future studies. Ordination analysis is employed in this study, such that species and samples are arranged in a low-dimensional space whereby similar sites are nearby and dissimilar entities far apart. Correspondence analysis ascertains the degree of ecological ‘‘correspondence’’ between sampling units and species using an eigen-analysis approach where sites and species arrangements are optimized by underlying latent variables (Thomas et al. 2002). An additional approach of canonical correspondence analysis (CCA) selects a linear combination of measured environmental variables to maximise the dispersion of sites and species. Both of these ordination techniques will be evaluated for determining the natural arrangement of sites. Clustering is an operation of multidimensional analysis which consists in partitioning the collection of objects in the study. Clustering techniques will be applied to ordination results to create classes that are based on species/cover composition and environmental variables. These classes can then be used to classify high resolution remotelysensed data. The combination of ordination, clustering, and image classification is important for analyzing imagery with very high spatial resolutions, where the potential exists for meaningful information to be derived from detailed ground information (Anderson & Clements, 2000; Jacobsen, Nielson, Ejrnaes, & Groom, 1999; Treitz, Howarth, & Suffling, 1992). This study will further the link between the theories and practices of classification of vegetation data by ecologists and image classification for mapping the spatial extent of vegetation by remote sensing scientists.
2
Study area
This study site is located at Cape Bounty on the south-central coast of Melville Island (74º55’N, 109º35’W) (figure 1). The study area at Cape Bounty is approximately 12km x 12km in size and is composed of two adjacent watersheds that drain into two separate lakes, and then south into Viscount Melville Sound. The area is underlain by steeply dipping sedimentary rocks of the Devonian Weatherall and Hecla Bay Formations and mantled with glacial and regressive early Holocene marine sediments (Hodgson et al. 1984). Continuous permafrost with an active layer of 0.5-1 m covers the entire study area. The climate is characterized by long, cold winters and a short, cool melt seasons from June-September. Mean daily July temperatures at Cape Bounty (2003 -2004) were 3.1°C and summer rainfall is infrequent, and typically of low intensity (