Habitat Classification Using Landsat 7ETM+ Imagery of the Ikh Nart
Total Page:16
File Type:pdf, Size:1020Kb
Mongolian Journal of Biological Sciences 2006 Vol. 4(2): 33-40 Habitat Classifi cation Using Landsat 7ETM+ Imagery of the Ikh Nart Nature Reserve and Surrounding Areas in Dornogobi and Dundgobi Aimags, Mongolia Daniel S. Jackson1, James D. Murdoch2 and Bayart Mandakh3 1Department of Biology, Oberlin College, Oberlin, Ohio 44074, USA* 2Wildlife Conservation Research Unit, University of Oxford, Tubney House, Abingdon Road, Tubney, Abingdon OX13 5QL, UK . 3Mongolian Academy of Sciences, Institute of Botany, Ulaanbaatar, Mongolia Abstract Wildlife studies increasingly employ thematic information, such as habitat classes, extracted from remotely sensed image data. Thematic information allows researchers to characterize biological processes of interest at various spatial scales. In and around Ikh Nart Nature Reserve of Mongolia, few large-scale habitat and landscape maps exist. Such maps are needed to support several long-term biological studies. Here we report the results of a maximum likelihood supervised classifi cation of a fi ve-band multispectral composite Landsat 7ETM+ image of the Dalanjargalan Soum section of the reserve and surrounding areas in Dornogobi and Dundgobi Aimags. We classifi ed the image into seven habitat classes: dense rock, low-density shrub, high-density shrub-rock mix, semi-shrub steppe, forb- dominated short grass steppe, tall vegetation, and ephemeral standing water bodies. Overall classifi cation accuracy was 90.5% with a Khat statistic of 88.8%, and a user’s accuracy of >85% per class, suggesting strong agreement between map and ground reference information. As such, the map presents a detailed distribution of habitats that may be suitable for a variety of research applications including the analysis of wildlife ranging behavior and the identifi cation of priority areas for conservation. Key words: Ikh Nart Nature Reserve, Landsat, map, Mongolia, supervised classifi cation, habitat, wildlife Introduction plan protected areas (Smith et al., 1997). As such, habitat classifi cation represents a valuable tool for Studies that describe biological processes at landscape conservation and research (Johnson et the local, regional, and global scale increasingly al., 2004). employ thematic information extracted from In Mongolia, satellite imagery has been used to remotely sensed satellite image data of the earth survey the country’s ecological landscape in several (Sampson & Delgiudice, 2006). Information studies. It has aided in linking seasonal patterns in from satellite imagery has been applied to diverse primary productivity with the migratory patterns contexts, from the quantifi cation of landscape of Mongolian gazelles (Procapra gutturosa, change (Lunetta et al., 2006) to the assessment of Leimgruber et al. 2001), and also been used to alternative energy viability (Dudhani et al., 2006). assess Mongolia’s vegetation cover (Naidansuren, In particular, habitat information from satellite 2003), and mitigate the growing impact of imagery has been applied to a variety of wildlife livestock (Rasmussen et al. 1999). Studies have studies (Hansen et al., 2001; Braun, 2005; Newton yet to employ high resolution imagery to explore Cross, 2007; Stevens et al., 2007). Johnson et al. the biogeographical characteristics of the Gobi- (2004), for example, used multispectral Landsat steppe ecosystem. images to map (and predict) suitable habitat for Ikh Nartiin Chuluu Nature Reserve in endangered mountain caribou (Rangifer tarandus Dornogobi Aimag (province) sits at the unique caribou) at multiple scales in British Columbia, confl uence of grassland and semi-desert steppe Canada. Other studies have used habitat zones of the Gobi-steppe ecosystem (Reading et information from satellite imagery to examine al., 2006) and is the site of several wildlife studies patterns of habitat use (Jepsen et al., 2002), monitor focusing on ungulates, carnivores, insectivores, variations in species richness (Gould, 2000), and and raptors (Reading et al., in press). The reserve 33 http://dx.doi.org/10.22353/mjbs.2006.04.13 34 Jackson et al. Ikh Nart Habitat Classifi cation is also currently in the initial stages of developing reserve along with a few freshwater springs, short a comprehensive, long-term, management plan (i.e., <50 m) streams, oases, alkaline pools, and (Reading et al., in press). To our knowledge, ephemeral ponds. The climate is arid and highly despite growing interest in the region’s fl ora and variable, with temperatures ranging from -40 °C fauna, no large-scale habitat or landscape feature in winter to +43 °C in summer. Precipitation is maps of the reserve exist (with the exception of low (<100 mm/year), falling mostly as rain from low resolution topographic maps). Accurate June through August. habitat maps would provide researchers with a standardized reference for studying wildlife Habitat classifi cation ecology and reserve managers with a tool to more Our classifi cation strategy involved 1) effectively plan conservation measures. identifying major, predominant habitat classes in We identifi ed major habitat classes in the the study area, 2) using known, ground-truthed northern section of Ikh Nart and mapped them sites to train a software classifi er on representative over a broad area by conducting a supervised pixels of a composite image (i.e., multiple overlaid classifi cation of a multispectral Landsat 7ETM+ bands), 3) classifying each pixel of the image into satellite image. Here we report the results of one of the habitat classes, and 4) evaluating the the classifi cation and describe the major habitat accuracy of the classifi cation by examining random classes. We aimed to generate an accurate and test sites on the ground. For the classifi cation, detailed habitat map suitable for use in current we used a Landsat 7ETM+ image (ID: 010-406: wildlife and conservation studies. WRS-2, Path 130, Row 028: eight spectral image bands; pixel resolution = 30 m x 30 m for bands Materials and Methods 1 to 5 and band 7) taken on July 26, 2002 from the Global Land Cover Facility (College Park, Study Area Maryland, USA). The Ikh Nart Nature Reserve (Ikh Nart) is a We identifi ed major habitat classes based on small protected area located in the northwest exploratory unsupervised classifi cations of the region of Dornogobi Aimag (N45.723° E108.645°; Figure 1). Established in 1996, the reserve protects ~66,000 ha of rocky outcrops and steppe habitats that harbor threatened argali sheep (Ovis ammon) and a myriad of other steppe fl ora and fauna (Myagmarsuren, 2000; Reading et al., 2006; Murdoch et al., 2006). Two soums (counties) jointly manage the reserve including Dalanjargalan and Airag. Our study area included the Dalanjargalan Soum section of the reserve (37,919 ha representing the northern 57%) and surrounding regions in Dornogobi and Dundgobi Aimags (total study area = 72,937 ha enclosed by N45.838943°- N45.545245°; E108.489732°- E108.731806°). The study area encompassed the reserve’s ‘core’ zone (7,120 ha; established to protect the argali sheep population) where several wildlife research projects are based. Ikh Nart typifi es the Gobi-steppe ecosystem, characterized by a mosaic of desert, semi-desert, and grassland habitats (Reading et al., 2006). Semi- arid vegetation dominates the reserve, including short grasses, shrubs, semi-shrubs, and forbs clustered among dense igneous and metamorphic Figure 1. Locaton of the Ikh Nart Nature Reserve, rock outcroppings (Reading et al., 2006). Several Mongolia relative to country (inset), aimag (black ephemeral river valleys and creek beds occur in the line), and soum (dashed line) borders. Mongolian Journal of Biological Sciences 2006 Vol. 4(2) 35 Landsat image, previous botanical surveys of We assessed the accuracy of the classifi cation Ikh Nart, and local knowledge of the area. We with fi eld test sites by generating a stratifi ed fi rst conducted unsupervised classifi cations with random sample of 60 test sites (pixels) per class MultiSpec software (version 3.0.2, Laboratory for (Fitzpatrick-Lins, 1981; Congalton & Green, Applications of Remote Sensing, Indiana, USA) 1999) using a Geographic Information System to classify various combinations of composite (ArcGIS v. 9.0, ESRI, Redlands, California, USA). images (three to fi ve bands) into fi ve to ten spectral Visiting each test site, we recorded the ‘actual’ classes (Jensen, 2005). We then referenced the habitat class of the site and the eight surrounding results with on-the-ground knowledge of the pixels (Jensen, 2005). major vegetation and landscape features of the We based classifi cation accuracy on overall, area and previous botanical surveys. We used user’s, and producer’s accuracy measures, and a this initial exploratory information to estimate target Coeffi cient of Agreement, or Khat statistic the number of distinct habitat types that would be (Jensen, 2005). Producer’s accuracy describes potentially resolved by the images in a supervised the probability of a correctly mapped reference classifi cation, but not to classify the fi nal habitat sample. User’s accuracy describes the probability image itself. that a sample from a classifi cation map actually For the habitat classifi cation, we performed a refl ects what is on the ground. The Khat statistic supervised classifi cation with MultiSpec software considers all class accuracies, providing a more using maximum likelihood methods and the comprehensive measure of map accuracy (see Decision Boundary Feature Extraction (DBFE) Jensen, 2005). We considered scores of at least 85% algorithm (Lee & Landgrebe,