Mapping of aggregated floodplain plant communities using image fusion of CASI and LiDAR data

Abstract

Combined optical and laser altimeter data offer the potential to map and monitor plant communities based on their spectral and structural characteristics. A problem unresolved is, however, that narrowly defined plant communities, i.e. plant communities at a low hierarchical level of classification in the Braun-Blanquet system, often cannot be linked directly to remote sensing data for vegetation mapping. We studied whether and how a floristic dataset can be aggregated into a few major discrete, mappable classes without substantial loss of ecological meaning. Multi-source airborne data (CASI and LiDAR) and floristic field data were collected for a floodplain along the river Waal in the Netherlands. Mapping results based on floristic similarity alone did not achieve highest levels of accuracy. Ordination of floristic data showed that terrain elevation and soil moisture were the main underlying environmental drivers shaping the floodplain vegetation, but grouping of plant communities based on their position in the ordination space is not always obvious. Combined ordination-based grouping with floristic similarity clustering led to syntaxonomically relevant aggregated plant assemblages and yielded highest mapping accuracies.

Introduction

Mapping is one of the most efficient methods to visualize trends of plant community (PC) patterns in space and time. The problem is that, although in most cases plant communities can clearly be recognized in the field, it is not always easy to detect the borders between different vegetation types. Vegetation units are abstractions of reality and boundaries between them are not always defined (Whittaker, 1973). Between two adjacent plant communities there is a transition that is either long or short (Sykora, 1984a, Sykora, 1984b). According to Barkman (1990) real boundaries may be absent (local continua) or sharp (usually man made), but both cases are rare and, in general, real boundaries are usually fuzzy. Glavac et al. (1992) state that in natural vegetation transitions between plant communities may be continuous or discontinuous, but usually the “non-linear but continuous” model (Scott, 1974) holds. In this spatial model, areas in which species composition changes very slowly are alternated by areas with quick species turn-over.

Core emphasis in natural vegetation mapping is on how to simplify vegetation gradients into discernible, mappable and ecologically meaningful units. Key challenges herewith are the delineation of classes with an adequate level of detail and solving the problem of transitional zones (Cingolani et al., 2004, Fortin et al., 2000). Remotely sensed reflectance data offer capabilities to map vegetation but require spectral signatures of pre-identified vegetation classes. Plant communities need to be defined prior to mapping. Various classification approaches were developed in the past years to cope with transitional zones, e.g. by means of fuzzy classifiers (Foody, 1996, Foody and Atkinson, 2002; Zhang and Foody, 2001). However, fuzzy approaches also need pre-identified classes as input and thus the problem of class delineation remains. Therefore, the emphasis of this work was on the development of a methodology that delineates plant communities that are discernible by RS data on one hand, and preserve maximal ecological significance on the other hand.

Plant ecologists have been developing methods to identify species assemblages for some time now (Hill, 1979), yet matching these methodologies with remote sensing (RS) techniques has only recently been tackled (Nilsen et al., 1999, Schmidtlein and Sassin, 2004, Thomas et al., 2003). In this respect, the use of multivariate statistical methods, such as cluster analysis and ordination, and ancillary abiotic field data is suggested to organize plant community data along environmental gradients across space. Such organization asks for the use of multispectral data of high spatial resolution to spectrally delineate such classes. Yet spectral discrimination of vegetation classes with similar species composition or environmental conditions remains subject to inaccuracies (Lawrence et al., 2006, Thomas et al., 2003).

In a previous study, multi-source remote sensing data were acquired over a semi-natural floodplain along the river Waal in the Netherlands (Geerling et al., 2007). Use was made of a compact airborne spectrographic imager (CASI) and light detection and ranging sensor technology (LiDAR). The authors concluded that a fusion of spectral and altimetry data can improve the classification of floodplain vegetation, particularly vegetation classes differentiated by their vertical structure. These classes are important to predict the hydraulic roughness of a floodplain, but were almost exclusively based on structure and not on species composition. As such, they lack syntaxonomic relevance.

Plant communities are usually described by plant ecologists at the lowest possible syntaxonomical level in the Braun-Blanquet system (Westhoff and van der Maarel, 1978). The Braun-Blanquet system describes plant assemblages at different hierarchical syntaxonomic levels with increasing number of floristic specifications: class, order, alliance and association. Differentiation of plant communities at lowest levels is based on the presence of indicator species, which are often to be found within the dominant plant cover. Since this level of differentiation is beyond the detection capacity of remote sensing instruments, the purpose of this work was to develop a methodology that enables to group these plant communities into broader vegetation classes in such a way that they can be discriminated by the CASI + LiDAR dataset and, at the same time, maintain phytosociological significance. Adopting ecological field data into the RS analysis can speed up the mapping process not only due to its potential use as ground truth data, but also by analyzing its inherent informative value for improving classification quality. Accordingly, the objective of this study was to evaluate the adequacy of implementing an ecological methodology using clustering and ordination techniques for optimizing floodplain vegetation mapping based on a fusion of spectral (CASI) and structural (LiDAR) data.

To do so, we first describe the phytosociological way to identify and describe plant communities, followed by a brief description of remote sensing as a means to identify and map plant communities. Subsequently we propose a synergism of both approaches by relying on plant community similarity analysis and ordination techniques to achieve an ecologically sound mapping.

Phytosociology identifies discrete plant assemblages by grouping species abundance data through cluster analysis and ordination (Austin and Smith, 1989, Fortin et al., 2000). Cluster analysis in field ecological studies essentially seeks to group floristic abundance data into classes or groups in such a way that within-group similarity is maximized and between-group similarity is minimized according to some objective criteria (Jongman et al., 1995). Basically, cluster analysis techniques divide plot samples of species (relevés) into a hierarchy of statistically similar clusters and then examine their dissimilarity. Amongst the most widely employed techniques are: two-way indicator species analysis (TWINSPAN) (Gauch and Whittaker, 1981, Hill, 1979) and cluster analysis by dendrogram method (McCune and Mefford, 1999). TWINSPAN is a polythetic, hierarchical clustering technique, which involves the joint clustering of samples and species by successive partitions of ordination axes generated at each step by reciprocal averaging. A dendrogram is a treelike plot that depicts the agglomeration sequence in cluster analysis, in which entities are enumerated along one axis and the dissimilarity level at which each fusion of clusters occurs is placed on the other axis. These methods show the hierarchy of clustering as a vegetation table showing all plots (TWINSPAN) and as a graphical representation (dendrogram), respectively. Due to differences in algorithm, the clusters made by TWINSPAN and dendrogram do not fully correspond. TWINSPAN classifies the plots in a divisive way, dividing the clusters using ordination space partitioning, whereas dendrogram groups the plots together based on a cluster distance measure in an agglomerative way.

Nevertheless, as vegetation units are abstractions of reality and boundaries between them are not always defined, in practice, clustering is often used in combination with ordination. Ordination analysis, as a kind of gradient analysis, seeks to detect a set of factors that account for the major patterns across all the original variables without a substantial loss of information. Detrended correspondence analysis (DCA; Hill and Gauch, 1980) is an ‘unconstrained’ ordination technique, which means that it refers to plot-based species data without considering any explanatory variables. A correspondence analysis (CA) is a form of weighted averaging that constructs theoretical environmental variables that best explain the species data (Jongman et al., 1995). This is done by maximizing the dispersion of the species scores along an ordination axis and the correlation between species and sites. The first axis symbolizes the ‘longest’ environmental gradient in species composition; the second one describes the ‘longest’ gradient in the remaining floristic variation and so on. Multiple axes can be constructed, with the constraint that they are uncorrelated with the previous axes. In a DCA two additional steps, detrending and rescaling, are added to remove two major faults: the arch effect that is caused by the unimodal distribution of species along gradients, and axis compression that occurs at the end of the gradients (McGarigal et al., 2000). Special problems arise with ordination in that it is subject to a number of assumptions about joint relationships of variables (Gauch, 1982).

In fact, despite case-specific shortcomings, ordination and cluster analysis may be seen as complementary, and when applied together they may provide useful information about the relationships among species (e.g. Thomas et al., 2003). In practice more than one solution is possible for defining plant communities. For instance, multiple plant assemblages can be defined, depending on the required level of detail, applied method, skills of the user and expert knowledge. Once plant assemblages are formed, they can be classified in syntaxa according to the Braun-Blanquet system. The philosophy of the Braun-Blanquet approach is that plant assemblages are biological units that can be typified and described into a hierarchical system in analogy to plant taxonomy (Westhoff and van der Maarel, 1978). The benefit of this classification system is that most of the described plant assemblages are quite well studied, particularly their synecology: the environmental conditions of plant communities.

Detection of plant assemblages through combined optical reflectance and laser altimeter data concerns the spectral and structural differentiation of a group of species, rather than the specific response of one species. Species assemblages have distinct spectral and structural characteristics when compared to single plants. The spectral influence of the non-vegetation elements, the multiple scattering between different plants and the structural layering of the vegetation can be seen as characteristic of a plant community (Schmidt and Skidmore, 2003). The success of PC mapping by optical remote sensing data depends therefore mainly on the type of remotely sensed data and the degree of syntaxonomical detail used for classification. Concerning the type of data, there are numerous examples of studies where high spatial resolution imaging spectroscopy data were applied to derive thematic maps of floristic composition (e.g. Schmidtlein and Sassin, 2004, Treitz and Howarth, 2000). Previous studies of river dynamics using multispectral and hyperspectral sensors include Winterbottom and Gilvear (1997), Malthus and George (1997) and Schmidt and Skidmore (2003). In addition, LiDAR data have significant potential for generating maps of complex river channel environments and vegetation characteristics. Airborne laser scanning is an active remote sensing technique that involves high-resolution elevation measurements by means of laser pulses. Over the past years various studies showed that laser data are able to distinguish structural differences both between and within plant communities (e.g. see review Straatsma and Middelkoop, 2006). For instance, Cobby et al. (2001), Mason et al. (2003) and Straatsma and Middelkoop (2007) employed laser altimetry for the assessment of floodplain vegetation height. The LiDAR observations of vegetation structure can thus present an independent information source complementing the spectral information content for a comprehensive floodplain characterization (Geerling et al., 2007, Gillespie et al., 2004, Held et al., 2003, Hill and Thomson, 2005).

Regardless of the type of remotely sensed data and the applied method, the syntaxonomical assignment of the ground truth data that precedes the classification step is usually taken for granted by remote sensing analysts. Given that various syntaxonomical assignments in the Braun-Blanquet system are possible from the same floristic data, it is this ‘grey zone’ of identification and, thus, the possibilities for improvement of mapping that will be further explored. The aforementioned methods of vegetation sampling, clustering and ordination will be used to delineate and optimize characterization of PC classes that can be discriminated through spectral and altimetry data. In Section 2 we will present a floristic dataset of a nature area that will be used for testing the map-ability of vegetation clusters.