Mapping vegetation in a heterogeneous mountain rangeland using landsat data: an alternative method to define and classify land-cover units

Abstract

Three major problems are faced when mapping natural vegetation with mid-resolution satellite images using conventional supervised classification techniques: defining the adequate hierarchical level for mapping; defining discrete land cover units discernible by the satellite; and selecting representative training sites. In order to solve these problems, we developed an approach based on the: (1) definition of ecologically meaningful units as mosaics or repetitive combinations of structural types, (2) utilization of spectral information (indirectly) to define the units, (3) exploration of two alternative methods to classify the units once they are defined: the traditional, Maximum Likelihood method, which was enhanced by analyzing objective ways of selecting the best training sites, and an alternative method using Discriminant Functions directly obtained from the statistical analysis of signatures. The study was carried out in a heterogeneous mountain rangeland in central Argentina using Landsat data and 251 field sampling sites. On the basis of our analysis combining terrain information (a matrix of 251 stands×14 land cover attributes) and satellite data (a matrix of 251 stands×8 bands), we defined 8 land cover units (mosaics of structural types) for mapping, emphasizing the structural types which had stronger effects on reflectance. The comparison through field validation of both methods for mapping units showed that classification based on Discriminant Functions produced better results than the traditional Maximum Likelihood method (accuracy of 86% vs. 78%).

Introduction

Ecosystems dominated by natural and semi-natural vegetation occupy large portions of the Earth's surface, and provide important ecosystem services that should be preserved (Balvanera et al., 2001). Such areas are generally destined to domestic grazing, which has been viewed as an activity with potential to meeting both goals of sustainable production and conservation Bloesch et al., 2002, Landsberg et al., 2003, Mohamed & Woldu, 2002. However, negative as well as positive effects of domestic grazing on biodiversity, primary productivity, and forage quality have been reported Milchunas & Lauenroth, 1993, Oesterheld et al., 1999, Perevolotsky & Seligman, 1998, West, 1993. Thus, a careful management-planning and further monitoring of rangelands become of main importance, while accurate base-line maps are indispensable for these purposes.

Landsat TM satellite images are a good tool for mapping vegetation (Jensen, 1996), although conventional supervised classification techniques have some inherent problems, due to differences in type and scale of information acquired by humans and satellites Cherrill et al., 1994, Keuchel et al., 2003, Wilkie & Finn, 1996. When the mapping area is complex and heterogeneous these problems are intensified, leading to mapping attempts of limited success (Budd, 1996). In rangeland ecosystems these difficulties are most likely to appear, because the influence of free ranging grazers combined to natural environmental gradients often create complex and heterogeneous vegetation patterns Adler et al., 2001, McIntyre et al., 2003. Three main types of such difficulties are faced when attempting to map rangeland ecosystems.

In the first place, there are problems originated by the limited spatial resolution of the TM satellite images. The main goal of traditional vegetation mapping has been the identification of plant communities (repetitive combination of species), or structural types (repetitive combination of growth forms and other terrain attributes) (e.g. Clark et al., 2001, McGraw & Tueller, 1983, Tobler et al., 2003, Wellens et al., 2000, Zak & Cabido, 2002). However, when communities or structural types are arranged in the landscape as patches smaller than the pixel size (30×30 m for TM images), attempts to map them are hampered (Clark et al., 2001). Training sites of adequate size may be impossible to find and, if found, the results of a supervised classification using those sites is inaccurate, especially if mixed pixels represent an important portion of the area. Therefore, a more realistic approach for mapping this type of landscape is needed, such as the definition of informational units (land-cover classes based on terrain attributes) at a higher hierarchical level, i.e. as combinations (mosaics) of communities or structural types (Davis et al., 1994).

The second problem is also related to the definition of informational units for mapping. Once the adequate hierarchical level is decided, the problem of defining discrete units discernible by the satellite still remains. When the basic components of the units to be defined (e.g. species, growth forms, community types) vary gradually, and to some extent independently, in response to multiple environmental and disturbance factors, the limits of the informational units for mapping must be imposed arbitrarily by the researcher Tanser & Palmer, 2000, Townsend, 2000, Wilkie & Finn, 1996, sometimes with the aid of multivariate classification techniques Cingolani et al., 1998, Coker, 2000, Jongman, 1987, Zak & Cabido, 2002. However, the basic components of the terrain selected by the researcher as variables for performing the classification may not be detected by the satellite (Millington & Alexander, 2000). This may lead to the definition of informational units that are meaningful for the researcher but cannot be discriminated by the satellite sensor, so producing inconsistencies due to different underlying approaches (Cherrill et al., 1994) and leading to a time consuming trial and error process until a satisfactory map is obtained (Clark et al., 2001). This being so, a preliminary analysis of the association between brightness sensed by the satellite and the various land-cover components perceived by the researcher would enhance the definition of land-cover units for mapping purposes (Armitage et al., 2000).

The third problem is related to the selection of the best training sites. Sometimes, training sites of adequate size for the defined land-cover informational units are difficult to find or recognize in the field. In such cases, several small training sites must be used to create spectral signatures defining a single unit Tobler et al., 2003, Wyatt, 2000. Depending on their characteristics, the various spectral signatures ought to be merged, maintained separately, or discarded as outliers, so leading again to a time-consuming trial and error process, until an acceptable set of signatures and an accurate final map are obtained. Even if large enough training sites for the different units could be obtained, the problem remains on how to select the most representative ones to perform the classification. Generally, the process is not straightforward, and an iterative and long procedure is the common rule to obtain acceptable results (Wilkie & Finn, 1996).

This paper addresses the way in which we have solved, using non-traditional approaches, these three types of problems obtaining an accurate map (based on Landsat imagery) of a heterogeneous mountain rangeland in central Argentina. Our approach was based on the following three points, each corresponding to one of the above-mentioned sets of problems: (1) definition of ecologically meaningful informational units as mosaics of different structural types; (2) consideration of brightness when defining informational units; (3) exploration of two alternative methods to perform the classification: the traditionally used Maximum Likelihood (ML) method, which was enhanced by analyzing objective ways of selecting the best training sites, and an alternative method using Discriminant Functions directly obtained from the statistical analysis of spectral signatures.

Considering these points, the objectives of this study were to: (1) define land-cover units useful for management purposes in the study area, based on structural attributes linked to the brightness data of Landsat ETM+ images, (2) explore objective methods for the selection of the best training sites, and perform a traditional supervised classification (ML) of the Landsat data, (3) perform an alternative method of classification taking maximum advantage of the spectral information of pixels in each of the spectral bands used, and (4) compare both classifications through field validation.